[Model] Add Sarvam MoE model support

[dsing-qual]

This PR adds support for the Sarvam MoE architecture in MLC LLM.

Changes include:

• Add sarvam_moe model implementation with:
  • Sarvam MoE config parsing
  • Qwen3-style attention with Q/K norm
  • Dense early MLP layers
  • Routed MoE expert layers
  • Shared expert branch
  • Sigmoid router with expert bias and routed scaling
• Add HuggingFace weight loading support for Sarvam checkpoints, including:
  • Embedding/name remapping
  • Packed QKV attention mapping
  • Packed gate/up MLP mapping
  • Routed expert weight stacking
  • Shared expert mapping
  • Router expert-bias mapping
• Register sarvam_moe in the model registry.
• This PR adds support for the Sarvam MoE architecture in MLC LLM.