Fix negative coordinates in tiled image sampling

[treeform]

Rotating a TiledImagePaint can produce negative source coordinates. The wrapped sampler uses signed mod, leaving negative pixel indexes that raise IndexDefect in debug builds and cause misplaced pixels or out-of-bounds reads in release builds. The example in #597 reproduces both symptoms on current master.

Wrap the base X/Y coordinates with an expanded integer template, following janAkali's diagnosis in the forum discussion. Image dimensions are positive, so a negative remainder can be corrected by adding the dimension. Derive the two adjacent indexes with boundary checks, reducing the sampler to two modulo operations without out-of-line wrapping calls.

Add regression tests comparing wrapped sampling with an explicitly repeated image, covering negative and fractional coordinates, tile seams, non-square and single-pixel dimensions, and five rotation angles including the reported -7 degrees. Include the tests in the standard runner and add tests/bench_images_tiled.nim for repeatable rendering/sampler benchmarks.

Fixes #597.

Validation on Windows with Nim 2.2.10:

• The new regression test fails on the original code and passes after the fix.
• Regression tests pass in debug and release builds.
• The original example passes in debug and release after the fix.
• The full standard test suite passes with both C and C++ backends.
• Existing tiled-image snapshots match exactly; git diff --check passes.

Performance on a Ryzen 9 5900X, Nim 2.2.10, Windows amd64, GCC 11.1.0, -d:release / -O3, no LTO: six interleaved rounds pinned to one logical CPU, with warmup and matching pixel checksums. Times are median milliseconds per operation. Shared comparisons use nonnegative source coordinates so the pre-fix build does not perform invalid reads.

Workload	Before fix	Initial function fix	This revision	Change vs before
Rotated 4x4 stripes, 256x160	1.185	1.521	1.210	+2.1%
Rotated 4x4 stripes, 1024x1024	30.529	39.234	31.482	+3.1%
Rotated 63x47 texture, 1024x1024	89.437	67.802	90.015	+0.6%
Unrotated 64x64 tiling, 1024x1024	13.757	25.756	12.500	-9.1%
Ordinary non-tiled smooth draw	23.191	14.586	23.146	-0.2%

Disassembly confirms the initial euclMod implementation leaves four external calls per wrapped sample; the revised sampler has none. The function variant also changes the compiler's placement of thread-local exception-flag lookups and happens to accelerate textured/non-tiled workloads. The revised version does not preserve every incidental gain of that variant; it stays close to the original baseline on the measured workloads. Results are specific to this compiler/platform.

Run the benchmark with nim c -r -d:release -d:benchAffinity tests/bench_images_tiled.nim. Append negative to include the original failing-coordinate case; that case must only be run against fixed builds.