Make AMD fp8 blockscale defaults hit the native fnuz MFMA fast path

[xzyaoi]

Finding

The native fp8 MFMA path is only selected by path="auto" when operands use an AMD-native fnuz fp8 dtype:

• Auto path predicate: src/xkernels/ops/gemm/triton/entry.py#L50-L57
• The quant helpers default to torch.float8_e4m3fn, not fnuz: src/xkernels/ops/gemm/reference.py#L137-L163 and src/xkernels/ops/gemm/reference.py#L166-L192
• The README notes the native MFMA path is the fast one and requires float8_e4m3fnuz operands; fn operands stay on the portable fallback.

Why this should improve performance

A caller using the public defaults can quantize to float8_e4m3fn and then call mm_fp8_blockscale(..., path="auto"), silently missing the 3-9x native fp8 MFMA path documented for MI300A. The optimized path exists, but the default API makes it easy not to use on AMD.

Suggested implementation

• Add a device-aware helper such as preferred_fp8_dtype() that returns float8_e4m3fnuz on AMD/gfx942 when available and float8_e4m3fn elsewhere.
• Consider a new fp8_dtype="auto" default for quant helpers, while preserving explicit dtype control for portability and parity tests.
• Add a warning or debug note when path="auto" falls back to portable because the operands are float8_e4m3fn on AMD.

Validation

• Add tests for dtype selection on environments with and without torch.float8_e4m3fnuz.
• Benchmark the default quantize + mm_fp8_blockscale(path="auto") flow before and after on MI300A.
• Keep explicit float8_e4m3fn tests to preserve portable fallback correctness.