add quantized qwen3 moe module

[TheoLee72]

Summary

Adds support for quantized Qwen3 MoE (Mixture of Experts) models to Candle. Implements efficient sparse MoE inference with top-k expert routing and weighted aggregation, supporting both 30B-A3B and 235B-A22B variants.

Changes

Core Model Implementation

• Sparse MoE architecture: Implements 128-expert MoE with top-8 routing per token using SparseMoeBlockWeights
• Weighted expert aggregation: Router computes softmax probabilities for expert selection, applies weighted sum with routing_weights
• Hybrid layers: Supports both MoE and dense MLP layers via MoeOrMlpWeights enum based on decoder_sparse_step
• GGUF metadata parsing: Extracts Qwen3-specific config (expert_count, expert_used_count, expert_feed_forward_length) from GGUF files
• Quantization support: Leverages existing QMatMul and indexed_moe_forward for efficient quantized inference

Technical Details

• Router precision: Converts inputs to F32 for router forward pass to ensure numerical stability in softmax
• Expert selection: Uses arg_sort_last_dim + narrow to extract top-k expert indices
• Runtime tuning: num_experts_per_tok can be overridden at inference time for speed/quality tradeoff

Motivation

Qwen3 MoE models offer superior efficiency compared to dense models by activating only 3.3B/22B parameters per token while maintaining 30B/235B total capacity. Quantized GGUF format (Q4_K_M) enables running these models on consumer hardware. This implementation fills a gap in Candle's model zoo for efficient large-scale MoE inference.

Breaking Changes

None - this is a new model addition with no modifications to existing APIs.
I only add "indexed_moe_forward" for QMatMul wrapper.

Validation

Functional Testing

30B Model (CPU):

cargo run --example quantized-qwen3-moe --release -- \
  --prompt "Write a Rust function to calculate the factorial of a given number."

235B Model (CUDA):

cargo run --example quantized-qwen3-moe --release --features cuda -- \
  --which 235b \
  --prompt "Explain the difference between MoE and dense architectures."

Expected output: Valid Qwen3-style responses with token/sec metrics printed.

---

References:

• Qwen3 Blog: https://qwenlm.github.io/blog/qwen3/
• Qwen3 MoE HF Docs: https://huggingface.co/docs/transformers/model_doc/qwen3_moe