Cache the x casts to fp8 and build ragged layout for fp8 gemm once

[MarioSieg]

Cache the ragged-group layout and x's fp8 cast in the grouped MoE FP8 GEMM path (fp8_grouped_gemm.py, moe.py) instead of rebuilding them from scratch on every w1/w2/w3 call — the layout and x are identical across w1/w3 (and the layout across w2 too), so this was pure duplicated work, verified bit-identical to the old behavior and measured ~15% faster on an isolated MoE-FFN forward at Qwen3-30B-A3B-ish dims. Companion fix to #3388's mxfp8 weight-quant caching, but for the plain (Hopper-compatible) FP8 path.

---

Note

Medium Risk
Changes the hot MoE FP8 forward path; correctness depends on callers only reusing layout/cache for matching (offsets, total_m) and tensor pairs, though the PR describes bit-identical outputs when used correctly.

Overview
Adds optional caching to the Hopper-compatible FP8 grouped GEMM so MoE forwards do not rebuild the ragged-group layout or re-cast the same activations on every w1/w3 (and layout-only reuse on w2 where x changes).

grouped_fp8_gemm and _GroupedFP8Gemm now accept optional layout and x_fp8_cache, with new helpers compute_grouped_layout and cast_grouped_input_to_fp8 for callers to precompute once. The gated and non-gated grouped-MM FP8 paths in moe.py compute layout (and x FP8 cast where w1/w3 share the same input) before the chained GEMMs and pass those into sibling calls.

Backward compatibility is preserved when the new arguments are omitted (same behavior as before).

^{Reviewed by Cursor Bugbot for commit 06b9a89. Bugbot is set up for automated code reviews on this repo. Configure here.}