perf(ROCm): parallel random sample with shared primitives

Row: GFX1100-TG200

Replaces the serial <<<n,1>>> random-sample kernel with a block-cooperative argmax reduction over Gumbel scores, mirroring the CUDA fix from #1984. The RNG and argmax reduce come from include/vt/sample_common.h, shared across CPU/CUDA/ROCm so bit-identity is a build property, not a copy-sync property. The serial path is retained behind VT_FAST_RANDOM_SAMPLE=0 for same-binary A/B gating.

Adds ROCm + CUDA subprocess A/B test cases that re-exec the binary with VT_FAST_RANDOM_SAMPLE=0 and =1 and assert byte-identical token ids across widths up to 248320 (Qwen3.8-27B vocab).

Verified: test_sampler 21/21 pass, 114 assertions. Token-identical to upstream baseline on Qwen3.5-4B Q4_K.

Start with include/vt/sample_common.h and locate the random-sample kernel and test_sampler entry points. Compare the serial and block-cooperative paths, then run the ROCm and CUDA subprocess A/B cases with VT_FAST_RANDOM_SAMPLE=0 and =1. Done means test_sampler passes and token IDs remain byte-identical across the stated vocabulary widths.