feat(xla): support Qwen2.5-VL windowed vision attention

[inureyes]

Parent and dependency

• Parent epic: epic: OpenXLA multimodal / VLM architecture support (Qwen2-VL, Gemma3n, Phi4MM, Molmo, and more) #566
• Blocked by feat(xla): support Qwen2-VL as the representative M-RoPE VLM #865
• Existing MLX implementation: src/vision/qwen2_5_vl.rs and src/vision/encoders/qwen2_5_vl.rs

Problem

Qwen2.5-VL is not a checkpoint-name alias for Qwen2-VL. Its vision tower uses RMSNorm/SwiGLU variants and alternates configured full-attention blocks with windowed blocks. The implementation computes a window_index, window-specific cumulative sequence lengths, and a restoration order around the patch merger. Reusing Qwen2-VL's packed full-attention graph would be numerically wrong even when tensor shapes load successfully.

Goal

Extend the qualified Qwen2-VL XLA path to Qwen2.5-VL with exact window partitioning/reordering, vision-block semantics, processor behavior, M-RoPE state, and end-to-end validation.

Required implementation

• Share the Qwen family processor, patch embedding, vision RoPE, artifact metadata, and M-RoPE language path where the existing MLX implementations are truly identical. Keep family-specific code/config explicit where semantics differ.
• Emit Qwen2.5-VL VisionRMSNorm, QKV attention, SwiGLU MLP, configured full-attention blocks, and window-attention blocks through StableHLO/IREE.
• Reproduce window_index, cu_window_seqlens, padding, spatial-merge-unit grouping, and inverse/restoration ordering from qwen2_5_vl.rs. Treat these as processor/graph inputs with validated static maxima or emit deterministic graph constants when fully shape-derived.
• Ensure no token from one window/media segment attends outside its permitted range. Full-attention block indices must come from checkpoint configuration and be included in artifact compatibility identity.
• Validate patch/temporal/merge/window sizes, block-index range and uniqueness, grid divisibility, reordered token count, and merger output count before native execution.
• Integrate Qwen2.5-VL family detection and filtered weight loading without falling back to the Qwen2 vision graph on unknown/missing configuration.
• Preserve the feat(xla): support Qwen2-VL as the representative M-RoPE VLM #865 CLI/server/cancellation/accounting path and feat(xla): support multimodal 3D RoPE positions and decode deltas #864 per-slot M-RoPE state.
• Keep video support explicit: validate at least one temporal-grid fixture if advertised; otherwise reject it with a family-specific unsupported error.

Non-goals

• Qwen3-VL fast position interpolation or deepstack feature injection.
• General dynamic-shape window attention.
• Reworking the already-qualified Qwen2-VL full-attention implementation.

Validation

• Pin a Qwen2.5-VL checkpoint/revision and compare processor tensors, window indices, both cumulative-length forms, reordered patch embeddings, a windowed block, a full-attention block, restored/merged features, M-RoPE positions, logits/KV, and greedy tokens against the MLX path and an independent oracle.
• Include fixtures where the grid spans multiple windows, contains padding at a window boundary, uses multiple media segments, and exercises every configured full-attention block.
• Add a negative test that intentionally uses Qwen2 full attention and proves the intermediate oracle catches the divergence.
• Run CLI and mixed continuous-batch server tests, including cancellation and slot reuse.
• Execute on a production-relevant IREE target and run all feat(xla): support Qwen2-VL as the representative M-RoPE VLM #865/Qwen2/text-only regression gates.

Acceptance criteria

• Window partition, cumulative lengths, permutation, and restoration match the existing MLX implementation exactly.
• Full/window block selection is checkpoint-driven and artifact-fingerprinted.
• Intermediate vision values and deterministic output match the pinned oracle within the documented stage tolerances/token exactness.
• Multi-window and multi-media attention isolation is tested.
• Image and advertised video behavior are either validated or explicitly rejected.
• Qwen2-VL and text-only regressions remain green.