[TRITON] Add MXFP4 quantization support to triton unified attention kernel

[amd-xiaoyu12]

Motivation

add MXFP4 quantization support to unified attention kernel

- Add Q-MXFP4, QK-MXFP4, and PV-MXFP4 quantization modes (0-3)
- Automatic fallback for incompatible HEAD_SIZE_PADDED
- Compatibility check: requires HEAD_SIZE_PADDED >= 32 and divisible by 32
- Support smoothed quantization with mean subtraction for better accuracy
- Add comprehensive tests

Files modified:
- aiter/ops/triton/_triton_kernels/attention/unified_attention.py (+184 lines)
- aiter/ops/triton/attention/unified_attention.py (+24 lines)

Files added:
- op_tests/triton_tests/attention/test_unified_attention_mxfp4.py (comprehensive test suite)
- op_tests/op_benchmarks/triton/bench_mxfp4_attention.py" (usage example with benchmarks)

MXFP4 modes:
- Mode 0: Original (no quantization, baseline)
- Mode 1: Native MXFP4 QK
- Mode 2: Smoothed MXFP4 QK (recommended)
- Mode 3+: Smoothed MXFP4 QK + PV

Usage: Set MXFP4_OPTION env var or pass use_native_fp4 parameter

Test Plan

Evaluated with Llama3 8B, QWen 3 32B and QWen 3 think 30B using gsk8k

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.