feat(gguf): add Qwen3.5 (qwen3-next) hybrid MoE GGUF loader

[emanueleDiVizio]

Summary

This PR adds full GGUF support for the Qwen3.5 architecture (internally qwen3-next), including hybrid GDN (GatedDeltaNet) + attention layers and MoE variants.

Qwen3.5 is not a standard transformer: it combines recurrent GDN layers, attention blocks, and mixture-of-experts routing. Supporting it in GGUF required implementing both the architecture and its tensor layout/dtype semantics.

This enables inference of quantized Qwen3.5 models (e.g. Qwen3.5-35B-Instruct GGUF) on both Metal and CUDA backends.

Key changes

Model support

• New quantized_qwen3_next.rs implementing:
  • Hybrid GDN + attention execution
  • MoE (including shared expert variant)
  • Dense (non-MoE) variant

GGUF compatibility fixes

• Correct conv1d weight layout transformation (kernel, dim → dim, 1, kernel)
• RoPE computation uses model dtype instead of forced F32
• QRmsNorm casts weights to input dtype (BF16 compatibility)
• V-head expansion matches GGUF tensor layout
• SharedExpert gate properly dequantized and reshaped

Runtime state

• Hybrid cache:
  • GDN recurrent state (fixed size)
  • Attention KV cache (sequence-dependent)
• Proper reset between requests

Testing

• Verified on:
  • Qwen3.5-35B-Instruct GGUF (MoE)
  • Qwen3.5 dense GGUF variant
• Tested on Metal and CUDA backends
• Validated multi-turn conversation correctness