Enable OCP FP8 scaled-mm path on gfx12

[TashaSkyUp]

Summary

• distinguish ROCm FP8 scaled-mm modes between OCP E4M3 and FNUZ
• ignore generic rocminfo aliases such as gfx12-generic when checking concrete gfx targets
• keep gfx950/gfx12 on the OCP E4M3 path instead of converting to FNUZ
• add FP8 pertensor tests for gfx1201 detection and ROCm mode-specific dtype behavior

Root cause

ROCm 7.2 reports gfx1201 together with the generic alias gfx12-generic. The previous regex parsed both as numeric gfx IDs, including 12, so Quark disabled torch._scaled_mm because 12 < 940. Forcing the old HIP path also converted OCP float8_e4m3fn tensors to float8_e4m3fnuz, which is not supported on gfx1201 in local testing.

Validation

• pytest -q test/test_for_torch/test_fp8_pertensor_kernel.py
• Local gfx1201 runtime validation outside this PR: Quark QParamsLinear uses torch._scaled_mm with torch.float8_e4m3fn and returns finite BF16 output on AMD Radeon AI PRO R9700.
• Local negative validation: forced hip_fnuz path passes float8_e4m3fnuz to torch._scaled_mm and fails with HIPBLAS_STATUS_NOT_SUPPORTED on gfx1201.