refactor(nvfp4): collapse load-time scratch into a single Model map

[kekzl]

Summary

Final cleanup of the NVFP4-prequant load path. Stage E (PR #72)
moved the runtime hot-path off the per-layer NvFP4PreQuantWeight
slots — dispatch reads from Tensor.scales / Tensor.tensor_scale
directly. The slots themselves remained as load-time scratch only
and were the last residual scaffolding in the type system.

This PR removes them.

Changes

• 13 NvFP4PreQuantWeight slots on TransformerLayer deleted
  (nvfp4_q/k/v/o, nvfp4_gate/up/down, nvfp4_w_*_shared)
• 3 std::vector<NvFP4PreQuantWeight> per-expert slots deleted
  (expert_nvfp4_gate/up/down)
• Model::nvfp4_out_proj_ deleted
• NvFP4PreQuantWeight moved out of TransformerLayer to a free
  struct in model_config.h (loses now-unused weight field;
  valid() checks weight_scale.data only)
• New single Model::nvfp4_scratch_ —
  std::unordered_map<std::string, NvFP4PreQuantWeight> keyed by
  canonical slot name:
  • "L{idx}.{slot}" for per-layer ("L5.wq", "L5.w_gate_shared")
  • "L{idx}.expert_w_{kind}.{e}" for per-expert
  • "out_proj" for LM head
• weight_map.cpp routes every NVFP4 scale into the map via a new
  route_nvfp4_scale(model, key, kind, t) helper. Replaces 4
  assign lambdas that wrote into the per-slot structs.
• weight_upload.cu iterates the map (single loop, ~30 LoC) instead
  of walking 16 per-layer slots × 3 fields each
• executor_pre_dequant.cu Phase 0 walks the map, resolve() parses
  the key back to the corresponding main weight tensor, promote()
  copies the device pointers + the FP32 tensor scalar onto its
  sidecars. Scratch is clear()ed at end.
• safetensors_loader.cpp drops the old "link" step entirely — it
  existed only to set NvFP4PreQuantWeight.weight so .valid()
  could return true; that field is gone

Architecture impact

TransformerLayer is now a plain description of the runtime weights
with no load-time scaffolding mixed in. Adding a new NVFP4 quant
variant (W4A16, future llm-compressor revisions, etc.) only touches
weight_map.cpp (routing) and executor_pre_dequant.cu (resolve
key → tensor). No more per-projection per-slot field churn.

Net diff: 6 files, +202 / -213 (~11 LoC reduction, -16 fields,
-3 vectors, -1 nested struct).

Test plan

• make test-gpu — 575 PASSED across 7 suites (1 pre-existing
  fail unchanged).
• LlmCompressorE2E.MistralSmall_LoadsAndGeneratesCoherent
  (llm-compressor reciprocal scaling path) — PASSED
• LlmCompressorE2E.Modelopt_QwenCoder30B_StillWorks
  (Modelopt multiplicative scaling path) — PASSED
• Qwen3-4B-Q8_0 GGUF smoke — 76 tok/s coherent (proves no
  regression on the non-NVFP4 path)

🤖 Generated with Claude Code