Add MXFP8 grouped GEMM (E4M3) for routed-expert MoE training

[ysa2215]

Description:

Why

GPT-OSS-style MoE training needs a grouped GEMM to run on MXFP8 — without it the MXFP8 dispatch path errored out on routed-expert layers, restricting MXFP8 to Linear targets. This is the minimum viable kernel to get MoE training running.

V1 uses E4M3 for all operands across fwd/dgrad/wgrad. A single format keeps the kernel free of dtype dispatch and the autograd free of a separate grad_output quantization path — simplest to implement and validate. The known cost is gradient
numerical robustness (E4M3's narrow dynamic range can underflow the long tail / overflow spikes of grad_output); the toy-MoE training test validates whether V1 is adequate in range.

The mixed-format recipe (grad_output in E5M2 — the TE / Meta / TorchAO industry consensus) is reserved for V2. To make that upgrade kernel-free, the API already exposes independent fwd_format / bwd_grad_format params (both defaulting to
E4M3) and the kernel carries a pre-written E5M2 branch; V2 just flips bwd_grad_format to E5M2. See MXFP8_GROUPED_GEMM_PLAN.md` §0 for the rationale.

What changed

Adds an MXFP8 contiguous grouped GEMM kernel (forward + dgrad + wgrad) and wires it into the MXFP8 dispatch path so routed-expert MoE layers can run on MXFP8.

• New kernel package alto/kernels/mxfp8/mxfp8_grouped_gemm/
  • cg_forward.py — Triton persistent forward grouped GEMM with super-grouping schedule and contiguous index routing, adapted from the MXFP4 scaffold with all K-packing removed (MXFP8 is one element per byte).
  • cg_backward.py — full backward (dgrad + wgrad) wrapped in an MXFP8GroupedGEMM(autograd.Function).
  • functional.py — user-facing mxfp8_grouped_gemm(...) plus the dispatch-layer entry _quantize_then_mxfp8_scaled_grouped_mm(...), which mirrors the MXFP4
    _quantize_then_mxfp_scaled_grouped_mm contract (1-D cumulative offs,
    padded activation buffers, weights kept in [E, K, N] dispatch convention to
    avoid a transpose copy).
  • autotune.py, __init__.py.
• Dispatch wiring (alto/kernels/dispatch/tensor.py): the MXFP8 weight wrapper's _grouped_mm path previously raised NotImplementedError; it now routes 2d-activation × 3d-weight + offsets calls to the new kernel, with guards limiting
  V1 to mxfp8_e4m3 and rejecting Hadamard/DGE/bias.
• Tests (tests/unittest/mxfp8/): grouped GEMM correctness with SNR gates, use_2dblock/use_dot_scaled coverage, cross-format comparison, and a toy-MoE end-to-end training sanity check (+ loss-curve plotting).
• Docs: MXFP8_GROUPED_GEMM_PLAN.md design/plan doc; m355 test results; README entry for MXFP8 (linear + grouped GEMM).

Test plan

• tests/unittest/mxfp8/test_mxfp8_grouped_gemm.py (SNR gates pass)
• tests/unittest/mxfp8/test_e2e_moe.py toy-MoE training sanity