feat(metal): fused MoE expert dispatch with Q4K kernels for Metal

[emanueleDiVizio]

Summary

This PR introduces a fused Metal execution path for Mixture-of-Experts (MoE) layers with Q4K quantized weights on Apple Silicon.

Instead of the standard pipeline:

route → gather → matmul → activation

this implements:

route → fused gather + quantized matmul + activation (single kernel)

This eliminates intermediate buffer materialization and reduces GPU dispatch overhead, significantly improving per-token latency for MoE models on Metal.

Key changes

Fused MoE kernel (Metal)

• New indexed_moe.metal kernel performing:
  • token routing by expert index
  • Q4K quantized matmul
  • fused gate + up projection with SiLU activation
• Uses 2 simdgroups (64 threads) tuned for Apple GPU execution

Runtime integration

• Automatically enabled when QTensor-backed weights are available on Metal
• Falls back to standard path when not available
• Supports batched expert matmuls in decode

Infrastructure

• Added moe_qtensor() access for direct QTensor buffer usage
• Eliminates redundant F32 conversions in Metal MoE path

Performance

Reduces per-token latency for MoE models (e.g. Mixtral Q4K) by:

• removing intermediate gather buffers
• reducing kernel dispatch count
• improving memory locality

Dependencies

Depends on:

• candle PR for QTensor Metal storage access (huggingface/candle#3444)