feat(mlx-lm): add Qwen3.5 hybrid attention model support

[Ogilthorp3]

Summary

Adds Qwen3.5 (hybrid Mamba + Transformer) model support to mlx-lm.

Qwen3.5 is a novel architecture mixing linear attention (Mamba-style SSM) with standard full attention (GQA + RoPE) in a 48:16 ratio across 64 layers. This is the first Rust implementation.

Changes

• **** (1,253 lines) — Full model implementation:
  • with nested and dispatch
  • : conv1d, SSM discretization (A_log, dt_bias), causal attention with B-scaled weighting, output gating
  • : GQA with partial RoPE (0.25 factor), q/k norms
  • enum dispatching between layer types
  • Quantized weight loading (4-bit affine, group_size=64)
  • Weight prefix stripping for VL-extracted checkpoints
• **** — Added
• **** — Added trait impl for qwen3_5

Tested with

• Qwen3.5-27B-4bit-text (64 layers, 5120 hidden, 248K vocab)
• Model loads in ~3s on M4 Pro (Mac Mini, 64 GB)
• Compiles with zero warnings

Notes

• SSM scan uses causal attention approximation — full exponential decay recurrence marked TODO
• All linear layers use for quantization support
• Follows existing qwen3.rs patterns exactly (derive macros, load functions, Module impl)

---

Built for Sanctum — a home AI platform running on Apple Silicon.