Add dynamic MXFP4 quantization kernel and tests

[suryajasper]

Implement a Wave kernel for dynamic MXFP4 (E2M1) quantization of f32 activations with E8M0 block scales, following the algorithm from ROCm/aiter's dynamic_mxfp4_quant.

The pipeline is split into three stages:

• Scale computation (compute_mxfp4_scales): per-32-element block amax via IEEE 754 bit manipulation, producing E8M0 scale bytes and a broadcasted quant_scale tensor. Runs as pure PyTorch on GPU since the 32-element reduction is incompatible with AMD wave64 (which requires >=64 elements for wave-level reduce).
• FP4 encoding (Wave kernel): comparison-based E2M1 encoding using seven threshold comparisons and select, entirely elementwise.
• Nibble packing (pack_mxfp4_codes): pairs of i8 codes into uint8.

Also adds f32_to_mxfp4 PyTorch reference in mxfp_utils.py for validation.

Tests:

• Lit test: FileCheck verifying key MLIR ops (absf, cmpf, select, fptosi) in the compiled kernel.
• Unit pytest: 4 parametrized shapes comparing Wave kernel output against f32_to_mxfp4 reference with dequantize round-trip.
• E2E pytest: Wave quant -> MXFP4 GEMM validated against the full PyTorch reference pipeline (f32_to_mxfp4 + torchScaledGemmMXFP4).