fix: invalid SPIR-V that fails HLSL legalization ([numthreads] on non-entry methods, implicit-LOD sampling outside pixel/compute)

[sasvdw]

PR Details

Summary

Two unrelated SDSL constructs emit invalid SPIR-V. On Direct3D11 both fail the same way.
LegalizeForHlsl throws a bare spvOptimizerRun InternalError, because nothing validates the module
first. This PR fixes both causes and reports the shared failure clearly.

1. [numthreads] on a non-entry compute method (#3303)

The compiler emitted OpExecutionMode LocalSize on the method that carries the attribute. The compute
entry point is a generated CSMain_Wrapper, so the mode landed on a non-entry function. The result was
invalid SPIR-V.

fxc and dxc accept [numthreads] on a non-entry function and then ignore it. The thread group size
comes only from the entry point. Stride now does the same.

• MethodOrMember.cs: emit LocalSize only when the attribute is on the compute entry
  (EntryPoint == ComputeShader). Otherwise skip it and log a warning that points to
  ThreadNumberX/Y/Z.
• ShaderMixer.cs: drain SymbolTable.Warnings to log.Warning in MergeSDSL. Before this change the
  code drained only the errors, so the warning never reached the user.
• ComputeEffectShader.cs: clarify the XML documentation. ThreadNumbers sets the thread group size of
  the compute shader, which is its [numthreads]. ThreadGroupCounts is the dispatch count.

This is not a regression. The classic 4.3 path also dropped a misplaced [numthreads].

2. Implicit-LOD texture sampling outside the pixel and compute stages (#3313)

A Texture.Sample(...) call reachable from a vertex, hull, domain or geometry entry point emitted
OpImageSampleImplicitLod. SPIR-V allows that instruction only in Fragment, GLCompute, MeshEXT
and TaskEXT. Implicit LOD needs screen-space derivatives, and those stages do not have them.

This change reports an error. It does not lower the call to SampleLevel(..., 0), because the shader
was never valid. I checked both Microsoft compilers directly:

	Texture0.Sample(s, uv) in a vertex shader
fxc /T vs_5_0	error X4532: cannot map expression to vs_5_0 instruction set
dxc -T vs_6_0	error: Opcode Sample not valid in shader model vs_6_0 → "Function uses features incompatible with the shader stage (vs)"
dxc -T vs_6_0 -spirv	error: sampling with implicit lod is only allowed in fragment and compute shaders
SPIR-V spec	ImplicitLod instructions require Fragment, GLCompute, MeshEXT or TaskEXT execution model

The Microsoft documentation for Sample agrees: "This method can be invoked within a pixel shader, but
it is not supported in a vertex shader or a geometry shader." The "Vertex-shader stage" page does list
Sample as legal in a vertex shader. That entry is a long-standing documentation error and should read
SampleLevel. SampleGrad and SampleCmpLevelZero do compile for vs_5_0.

So this is not a functional regression, because such a shader never compiled. It is a diagnostics
regression. Version 4.3 reported X4532. Version 4.4 reported spvOptimizerRun InternalError.

The engine has always treated this as illegal. That is why MaterialDisplacementMapFeature sets
IsNotPixelStage and routes texture reads through ComputeColorTextureLod, which uses SampleLevel.

• ShaderMixer.cs: after interface processing, walk the call graph from each non-fragment entry point.
  Report every implicit-LOD sample that the walk reaches, with its source location and the explicit-LOD
  alternatives:

  VSTextureSample.sdsl(14,25): A texture sample with an implicit level of detail needs a pixel or
  compute shader. A Vertex shader reaches this code. Use SampleLevel, SampleGrad or SampleCmpLevelZero
  to give an explicit level of detail.
  

A silent lowering to LOD 0 would make Stride accept a shader that fxc, dxc and the SPIR-V validator all
reject. It would also change the sampling behavior without telling the author. I am happy to change this
if you prefer the permissive behavior.

3. Shared diagnostics

• EffectCompiler.cs: when LegalizeForHlsl throws, validate the module and report the instruction
  that caused the failure, instead of a bare InternalError. Issue Texture.Sample() in a vertex-stage material surface emits invalid SPIR-V (D3D11 compile fails with a bare InternalError) #3313 asks for this. It also covers
  future problems of the same family.

Related Issue

Fixes #3303.
Fixes #3313.

Repro for #3303: https://github.com/sasvdw/stride-spirv-legalize-repro
Repro for #3313: https://github.com/delustra/stride-vtf-spirv-repro

Types of changes

• Docs change / refactoring / dependency upgrade
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist

• My change requires a change to the documentation.
• I have added tests to cover my changes.
• All new and existing tests passed.
• I have built and run the editor to try this change out.

Validation status

• Tests: StrideShaderTests gains TextureSampleOutsideFragmentStageIsReported, which uses
  CompilerTests/VSTextureSample.sdsl. It also gains TextureSampleInFragmentStageCompiles, which uses
  CompilerTests/PSTextureSample.sdsl. The second test is the control. It proves that pixel-stage
  sampling still compiles to valid SPIR-V. NumThreadsOnNonEntryMethodIsIgnoredWithWarning still passes.
  I checked each test red before the fix and green after it. The test for Texture.Sample() in a vertex-stage material surface emits invalid SPIR-V (D3D11 compile fails with a bare InternalError) #3313 is a separate commit that
  comes before its fix, so the failure is on record.
• Suite: 518 tests pass. 10 failures are pre-existing on this branch and unrelated. The test harness
  parses floats from .sdsl test headers with the current culture, so they fail on any comma-decimal
  locale. The same tests fail when I revert these changes. fix: parse and format shader numbers with the invariant culture #3325 fixes them separately.
• Editor: outstanding. This is the mandatory personal-testing step.
• Docs: the XML documentation for ComputeEffectShader.ThreadNumbers and ThreadGroupCounts is now
  clearer, and it flows to the API reference. The manual has no compute-shader guide. I track that as a
  separate stride-docs follow-up.

[github-actions]

🤖 Draft PR — automatic CI is skipped to save runner minutes.

• Mark the PR ready for review to run the full automatic CI — or add a ci-run-on-draft label to run it now without leaving draft.
• Or arm a specific opt-in suite: ci-enduser, ci-editor, ci-ios, ci-android.