feat(moe): serve NVFP4 and MXFP4 experts fused through the expert banks

[jason-fxz]

Resident (--moe-strategy fused) experts now load through the offload expert-bank loader instead of the state dict: HF pieces or FTW banks are packed by the kernel's pack() straight into per-layer GPU banks. Fused therefore serves bf16, fp8_block, NVFP4 and MXFP4 (gpt-oss) alike; GGUF q4_0 and DeepSeek-V4 experts stay offload-only.

• The per-format create_weights / resident_view overrides are gone; MoEMethod has one default for every format.
• On unified-memory GPUs (GB10) auto picks fused for these formats. Offload-only flags are ignored there with a warning; an explicit --moe-strategy offload still honors them.
• ft checkpoint always writes MoE experts as banks, since every strategy reads them.

Breaking: an FTW converted with --moe-backend fused or triton keeps its experts dense and no longer loads; reconvert it with ft checkpoint.

Follows up #445 and removes its _fused_resident_ok FIXME. Part of #369: this implements the resident NVFP4 / gpt-oss MXFP4 steps proposed in this comment. The sm_121 backend-probe audit and the install docs from #369 remain open.

Validation

H100 80GB, greedy, four bs=1 prompts plus one batched run, token ids compared with main:

• Offload family unchanged on 15 configs: Qwen3-30B, Qwen3.6 bf16 / FP8 / NVFP4, gpt-oss HF and FTW, Gemma-4 bf16 / NVFP4, truncated GLM-4.7 / MiniMax-M2.5 / DeepSeek-V4, plus cpu, hybrid and --moe-cpu-layers.
• Fused bf16 / fp16 / FP8 / gpt-oss identical to fused on main; gpt-oss FTW now serves fused (KeyError on main). The new NVFP4 fused (Qwen3.6, Gemma-4, MiniMax) is identical to offload on main.
• ft serve --moe-strategy fused with Qwen3.6-35B-A3B-NVFP4 answers a chat request correctly.
• tests/moe tests/engine tests/models tests/checkpoint tests/layers: no new failures versus main.

Not tested: Marlin / b12x resident, TP > 1, a real GB10.