perf(moe): split hybrid decode routes inside the ensure kernel

[jason-fxz]

Summary

After _ensure_experts_hybrid_kernel rewrites a layer's routes to cache slots, hybrid decode built the GPU and CPU inputs with torch ops on every MoE layer: a clone of the ids, a >= 0 compare, two torch.where, a clamp_min(0) for MoE kernels without supports_inactive_slots, and with collect_stats three sums and eight in-place adds in record_decode_stats_hybrid. The ensure kernel now does this in the same launch:

• It writes the CPU's routes (expert id, -1 for a GPU route) and the GPU's weights (0 for a CPU route) next to the GPU slots.
• CLAMP: CPU routes get slot 0 instead of -1 when the GPU kernel reads every route's slot.
• STATS: the hybrid missing / fetched / active / calls counters and their per-layer copies. The stat_* tensors become views of one buffer, so the kernel takes them as two pointers; decode_miss_stats and decode_miss_stats_per_layer read them as before.

ensure_experts_hybrid now takes the router weights and returns HybridRoutes(gpu_slots, gpu_w, cpu_ids). _decode_hybrid is its only caller, so the path without the split goes away: record_decode_stats_hybrid, the CPU reference _ensure_experts_hybrid_cpu, the two tests in tests/moe/test_hybrid_fetch.py that drove the kernel without the split, and the split-rule test there that only checked its own reference helper are removed.

This is independent of #670. Both change _decode_hybrid, in different hunks; a trial merge of the two branches merges without conflicts.

Validation

H100 80GB HBM3, Xeon Platinum 8480+, torch 2.14.1+cu130.

• Equivalence with main (a script, not added to tests/): the same 300 random decode steps (3 layers, 32 experts, 40 slots, top-8, batch 1 and 4, fetch fraction 0.5, clamp on and off, fp32 and bf16 weights) through main's torch split with record_decode_stats_hybrid and through HybridRoutes, in two processes. GPU slots, CPU ids, GPU weights, slot_for_id and id_of_slot are equal after every step; the totals and per-layer stats are equal at the end.
• tests/moe tests/engine/test_moe_cpu_layers.py on 930cc18: 155 passed, 6 skipped. Main gives 157 passed on this machine; the difference is the two removed tests. 47aa1a1 then drops one more test, which ran only Python code; the 2 tests left in tests/moe/test_hybrid_fetch.py pass.
• Decode throughput, bench_decode_moe.py --backend hybrid --greedy --decode 512 --cache-rate 0.25, ABBA against main (9c635a1), tok/s, measured on the first commit (13a4887):

Model	main	this PR
Qwen3.6-35B-A3B-FP8	106.60 / 106.72	111.37 / 111.09
nvidia/Qwen3.6-35B-A3B-NVFP4	126.86 / 126.93	132.66 / 132.64

Output text sha1 matches main in every run. On the final commit, one Qwen3.6-35B-A3B-FP8 hybrid run decodes at 111.08 tok/s with the same output sha1.

• End to end on a local merge of this PR and refactor(cpu-moe): split the executor into weight formats and gpu sync backends #670, Qwen3.6-35B-A3B-FP8 and nvidia/Qwen3.6-35B-A3B-NVFP4, 2560 expert slots (25%):
  • tests/e2e/test_aime.py with AIME25 problem 1 and the hybrid strategy passes on both models;
  • 4 AIME prompts decoded together (greedy, 512 tokens, in-process engine, CUDA graphs up to batch 4): with hybrid and moe_collect_stats on, the texts, decode_miss_stats and decode_miss_stats_per_layer equal main's; with the cpu strategy the texts equal main's;
  • served batch-1 decode (bench_decode_moe.py --backend cpu,hybrid --greedy --decode 512 --cache-rate 0.25): output sha1 matches main's for hybrid and 51afda3's (the base of refactor(cpu-moe): split the executor into weight formats and gpu sync backends #670) for cpu on both models.

Not covered: ROCm.