[CuTe] Validate dynamic tiler divisibility for static layouts (issue #3513)

[XFDG]

Summary

Add runtime shape-divisibility validation when layout composition combines a fully static left-hand layout with a dynamic right-hand tiler.

The existing logical_divide core test now covers both:

• a compatible dynamic tiler that preserves the input layout size
• the issue reproducer, which must fail with Shape Divisibility Condition

Rationale

The composition path already diagnoses fully static incompatibilities at compile time. When the right-hand shape is dynamic, however, new_shape and rest_shape become dynamic and the static assertion cannot run. The fallback runtime assertion was skipped even though each left-hand mode was still statically known, allowing an incompatible divide to silently return a non-size-preserving layout.

Limiting the runtime assertion to a statically known curr_shape closes that mixed static/dynamic gap without changing the intentionally permissive fully dynamic path.

Validation

Compiled the same standalone reproducer before and after this change with CUDA 13.1 and -arch=sm_100a, then ran it on an NVIDIA B200.

Before:

• dynamic tiler 128 returned normally
• input size was 384, but result size was 288
• the probe exited with status 2

After:

• the same invalid tiler aborts with Shape Divisibility Condition (status 134)
• a compatible dynamic tiler 12 returns normally with input/result size 384
• git diff --check passes

The cluster image does not provide GTest, so the added core regression test is left for the upstream test target to compile and execute.

Fixes #3513