fix(moe): match the benchbw verdict to the served model's expert size

[bernimccoy]

Fixes the 3x decode regression of --moe-backend auto on small-expert NVFP4 models
(Qwen3.6-35B-A3B-NVFP4 on an RTX 4070 SUPER: auto -> hybrid 27 tok/s, offload 84 tok/s).

Problem

load_backend_recommendation keys the profile on the expert format only and prefers the
ft bench bw --dtype verdict, measured on one geometry per format (nvfp4: 3072x1536, 7.6 MiB
experts). The CPU MoE kernel does not scale down to small experts: the same box benches
89.7 GB/s there and 32.2 GB/s at Qwen3.6-35B-A3B's 2048x512 (1.7 MiB), below the 2x PCIe
threshold. So auto picks hybrid and decodes 3x slower than offload. ft bench bw --model qwen3.6-moe already records offload for this geometry, but the reader consults the
per-format dtypes entry first and never reaches the per-model one.

Fix

load_backend_recommendation / load_hybrid_fetch_fraction take the served model's
geometry (H, I, E, top_k) and expert_bytes:

• a per-model workload entry benched at that geometry wins;
• otherwise the dtype entry applies only if its expert_bytes is within
  EXPERT_BYTES_TOLERANCE (2x) of the model's; other workload entries are filtered the same
  way; a skipped entry logs once with the ft bench bw --model remedy;
• callers that pass no geometry, and profiles written before expert_bytes was recorded,
  get today's format-only match.

The engine passes both at the backend pick and at --moe-hybrid-max-fetch auto. No profile
format change.

Tested on

RTX 4070 SUPER 12 GB (sm_89), driver 591.86, WSL2, torch 2.11.0+cu130, 24-core host, 108 GB
RAM. nvidia/Qwen3.6-35B-A3B-NVFP4, --moe-cache-auto (3831 slots), bs=1.

benchmarks/bench_decode_moe.py --backend offload,cpu,hybrid (256 decode tokens):

backend	decode tok/s
offload	83.7
cpu	25.9
hybrid (what auto picked on main)	27.2

ft bench bw: dtype nvfp4 89.7 GB/s CPU vs 25.3 PCIe (hybrid); --model qwen3.6-moe: 32.2
vs 25.9 (offload). With this change and the same profile, ft serve (auto) logs
benchbw profile: the 'nvfp4' entry was benched on 7.61 MB experts, this model's are 1.69 MB (4.5x apart); not applying its verdict, picks offload, and streams 91.0 tok/s on the request
that gave 26.4 tok/s under auto -> hybrid.

tests/moe/test_hybrid_fetch.py: new case with the measured profile numbers (exact-geometry
entry wins; a 4.5x-off dtype verdict is skipped; a 1.3x-off one still applies; no-geometry
callers unchanged); 6/6 pass. bf16 Qwen3-30B-A3B (9.4 MiB experts) is within 2x of the bf16
dtype geometry, so its verdict is unchanged.

Caveats

The 2x tolerance is a judgment call from one box (the kernel ratio here was 2.8x between the
two geometries); a per-model bench stays the precise answer and the warning points at it.

Assisted by: Claude Code