AML-4269: Add Location Support to Qwen3.5

[anvdn]

Attention Heatmap Support for Qwen3.5 + refactor

Summary

Extends attention heatmap capture to Qwen3.5 (hybrid linear/full-attention model) and refactors the existing heatmap code so any model can opt in with minimal boilerplate.

Changes

• New AttentionHeatmapQueryRecorderMixin (hs/attention_heatmap.py): centralizes query-buffer allocation and per-layer query recording. Qwen2-VL, Qwen3-VL, and Qwen3.5 now share this implementation instead of duplicating it.
• Qwen3.5 integration (models/qwen3_5.py): full-attention decoder layers now return q; linear-attention layers are skipped. The model registers itself via the mixin.
• Layer selection API change (server_args.py): replaced the attention_heatmap_layer_start / _layer_end range with a flexible attention_heatmap_layer_ids: list[int]. This is required for hybrid models where capturable layers aren't contiguous.
• Scheduler hybrid-pool support (scheduler_output_processor_mixin.py): when the KV cache is a HybridLinearKVPool, key tensors are looked up via full_attention_layer_id_mapping instead of raw layer id. The model now owns the canonical list of recorded layer ids.
• Version bump to 0.5.11+hs3.