Add initial NVIDIA GPU backend bring-up

[zhoubot]

Summary

This PR brings up an initial NVIDIA GPU backend for PTO Tile Lib, centered on DGX Spark / GB10 (sm121), and adds a real GPU test lane.

Included in this PR

• initial CUDA GPU backend scaffolding and dispatch wiring
• sm121 matmul fast-path groundwork
• tensor-core WMMA matmul fast paths for half / bf16 on sm121
• extended TMATMUL, TMATMUL_ACC, TMATMUL_BIAS, TMATMUL_MX, and TGEMV_MX GPU coverage
• standalone CUDA correctness test lane under tests/gpu/st
• lightweight GB10 matmul microbenchmark
• GPU-specific swizzle tile layout (SLayout::GpuSwizzle128B) that is intentionally separate from NPU boxed layouts
• larger 64x64x64 GEMM correctness tests for half / bf16

Notes

• current float matmul still uses an inline-PTX FMA fallback path
• current MX wrappers accept the scale tiles but reuse the existing GPU matmul path; scale semantics are not fully modeled yet
• current GPU swizzle layout is groundwork for future shared-memory / tensor-core-friendly paths and is not yet consumed by the sm121 matmul fast path

Validation

Executed on the target GB10 / DGX Spark environment:

• cmake --build build/tests/gpu-st -j4
• ctest --output-on-failure
• ./build/tests/gpu-st/testcase/pto_gpu_perf/pto_gpu_perf

Representative benchmark signal from GB10 (64x64x64, 1 block):

• float: ~2.2036 ms, ~0.24 GFLOPS
• half: ~0.0082 ms, ~63.86 GFLOPS
• bf16: ~0.0082 ms, ~64.03 GFLOPS