Add Qwen3.5 model implementation - #6

Summary

• Implements the Qwen3.5 text decoder with hybrid linear/full attention (Gated Delta Net)
• Supports loading pretrained weights from Qwen/Qwen3.5-0.8B
• Follows the same single-file, self-contained pattern as other models in minhub

Architecture highlights

• Hybrid attention: 18 linear attention layers (Gated Delta Net) + 6 full attention layers (every 4th)
• Full attention: GQA with output sigmoid gate, QK RMSNorm, partial RoPE (25% of head_dim)
• Linear attention: Causal conv1d, recurrent gated delta rule with learned decay (A_log) and step size (dt_bias)
• On-the-fly computation: Causal masks and RoPE frequencies computed per-forward to support 262K context without large buffer allocation

Validation

• Logits match HuggingFace transformers reference (Python, float32) to ~1e-3 tolerance

Test plan

• Custom small config forward pass (runs on CI)
• Pretrained model logit comparison against Python reference (skip_on_ci)