refactor(musa): remove direct flaggems calls from vendor backend

[physics31415926]

Summary

Refactor MUSA backend to remove direct FlagGems calls for topk_softmax. This method is now dispatched through the standard op manager system.

Changes

• Removed topk_softmax method from MusaBackend class
• topk_softmax now routes through op manager dispatch system, using FlagGems implementation when available

Testing

Tested on MUSA hardware (moer_15 machine) with MoE model inference:

Environment:

• Hardware: MTT S5000 (8x 80GB GPUs)
• Driver: 3.3.5-server
• Container: vllm_graph_torch29_new
• Model: Qwen3.6-35B-A3B (67GB MoE model)

Test Configuration:

• Tensor Parallel: 4 (GPU 2-5)
• Max model length: 2048
• Mode: Eager execution
• GPU memory utilization: 0.9

Result: ✅ PASSED

• Model loaded successfully (67GB in 15.5s)
• topk_softmax correctly dispatched to FlagGems: Patched vllm_topk_softmax for MUSA (FlagGems) in: fused_topk_router, fused_topk_bias_router
• Inference successful with correct outputs:
  • "Hello, my name is" → "John. I am a 25-year-old male, and I have been experiencing a persistent cough"
  • "The capital of France is" → "Paris, a city renowned for its iconic landmarks such as the Eiffel Tower, the Louvre Museum"

Performance:

• Throughput: ~3.6 toks/s output, ~0.9 toks/s input
• KV cache: 2.2M tokens allocated
• Max concurrency: 1086x for 2048-token requests