perf(moe): speed up grouped MoE routing

[Glitchfix]

Summary

This PR tightens the routing/unrouting path in Qwen3VLMoeTextExpertsGroupedMm:

• keep routed token indices as a 1D vector instead of expanding them across the hidden dimension
• use index_select / index_add_ for route and combine instead of hidden-size-expanded gather / scatter_add_
• update the standalone MoE comparison harness so it calls the current expert API and checks grouped output against the PyTorch reference

The old path was doing a lot of unnecessary index traffic. That is mostly invisible for smaller routing configs, but it becomes very noticeable when top_k and the expert count go up.

Benchmarks

Environment: RTX 6000 Ada, torch==2.10.0+cu128, BF16 unless noted.

Focused MoE harness:

python -m cosmos_framework.model.vfm.vlm.qwen3_vl_moe.moe_test

naive:      13.81 ms
optimized:   3.92 ms
speedup:     3.53x
relative diff: 0.00531

Synthetic decoder stack throughput, including attention + norms + MoE, old grouped routing vs this PR:

E128_K8_I768, 8 layers, seq=1024
old: 134.58 ms,  7,608.7 tok/s
new:  42.16 ms, 24,289.3 tok/s
speedup: 3.19x, +219.2%

E128_K8_I768, 12 layers, seq=1024
old: 136.89 ms,  7,480.3 tok/s
new:  63.33 ms, 16,170.2 tok/s
speedup: 2.16x, +116.2%

E128_K8_I1408, 8 layers, seq=1024
old: 103.66 ms,  9,878.7 tok/s
new:  52.81 ms, 19,391.2 tok/s
speedup: 1.96x, +96.3%

For the default-ish E=60, top_k=4 shape, I did not see a meaningful total-throughput improvement. The gain scales with routing pressure; high-expert / high-top_k configurations are where the old expanded-index route/combine path becomes expensive.

Tests

LD_LIBRARY_PATH= uv run --no-sync python -m ruff check \
  cosmos_framework/model/vfm/vlm/qwen3_vl_moe/moe.py \
  cosmos_framework/model/vfm/vlm/qwen3_vl_moe/moe_test.py

LD_LIBRARY_PATH= uv run --no-sync python -m cosmos_framework.model.vfm.vlm.qwen3_vl_moe.moe_test

LD_LIBRARY_PATH= uv run --no-sync python -m cosmos_framework.model.vfm.vlm.qwen3_vl_moe.moe_bench \
  --compare --num-tokens 256 --num-iters 10 --num-warmup 3 --no-compile --dtype fp32

LD_LIBRARY_PATH= uv run --no-sync python -m cosmos_framework.model.vfm.vlm.qwen3_vl_moe.moe_bench \
  --compare --num-tokens 2048 --num-iters 20 --num-warmup 5 --no-compile --dtype bf16

LD_LIBRARY_PATH= uv run --no-sync python -m cosmos_framework.model.vfm.vlm.qwen3_vl_moe.moe_bench \
  --num-tokens 1024 --num-iters 5 --num-warmup 2 --dtype bf16 --backward

[lfengad]

overall LGTM