kernels: support AMD-FP4 data type and low-precision training

[zhitwang17]

Add AMD-FP4 low-precision training support

Summary

This PR introduces AMD-FP4 as a first-class FP4 quantization scheme alongside
the existing NVFP4 path. AMD-FP4 shares the NVFP4 micro-block layout (E2M1 elements +
FP32 per-tensor outer scale) but uses a UE5M3 inner-block scale (GFXIPARCH-2067
§19.10 / OCP E5M3) instead of E4M3, giving ~256× wider inner-scale dynamic range at the
same block size. The kernel stack is refactored so NVFP4 and AMD-FP4 are peer recipes
that reuse a single block-quant implementation, differing only in the inner-scale grid.

Changes

Kernel reorganization

• Renamed fp4/fp4_common → fp4/fp4_primitives (moved grouped_utils,
  tensor_wrappers, triton_fp4_ops) and introduced a shared
  fp4/outer_scaled_fp4/ layout (api, kernels, pack_unpack, scales) that backs
  both NVFP4 and AMD-FP4. NVFP4 quantization shrinks ~720 lines by delegating to the
  shared primitives.

Inner-scale dtype primitives

• Added e4m3_ops and ue5m3_ops inner-scale dtype primitives, with ue5m3 providing
  the wider dynamic range required by AMD-FP4.

scale_format-aware quantization

• convert_to/from_* now take a scale_format ("e4m3" | "ue5m3") so a single
  block-quant routine serves both schemes. The legacy
  convert_to_nvfp4(scale_format="ue5m3") path is deprecated in favor of the dedicated
  AMD-FP4 ops.

AMD-FP4 ops + dispatch routing

• New alto::convert_to_amdfp4 / convert_from_amdfp4 ATen ops (inner scale pinned to
  ue5m3), plus amdfp_linear and amdfp_grouped_gemm.
• Dispatch config adds "amdfp4" as a precision peer of "nvfp4" with an
  inner_scale_format field; __post_init__ forces ue5m3 for amdfp4 and rejects
  mismatched configs. tensor/conversion updated to route AMD-FP4 calls.

LPT modifier

• LowPrecisionTrainingModifier accepts "amdfp4" as a valid scheme and applies the
  same lora_rank % 16 == 0 constraint as NVFP4.

Test status

Op-Level Test

All AMD-FP4 and NVFP4 unit tests pass: 748 passed, 0 failed (~111s) on MI300X / ROCm.

• New tests/unittest/amdfp4/ suite: quantization, linear, grouped GEMM, UE5M3/E4M3
  dtype, dispatch guards, Triton↔PyTorch parity, and an A/B matrix cross-checking
  AMD-FP4 against the reference path.
• Extended tests/unittest/nvfp4/ suites for the new scale_format parameter and shared
  primitives (no regressions).

E2E Test

Debug Model

GPT-OSS debug model, VS BF16 and NVFP4

E2E full test with GPT-OSS-20B