Phase 2: Paged decode attention over real block tables

[inureyes]

Context

The decode kernel currently slices per-sequence dense buffers using an identity block table built by PagedDecodeMetadata::from_visible_lengths ([0, 1, 2, ...]). With real pool storage in place, attention must gather the sequence's actual, possibly non-contiguous, physical blocks.

Phase 0 outcome (ADR 0001, #117): option A is gather-then-SDPA — take(pool, block_ids, axis=0) (layout A), then reshape, then transpose [0,2,1,3], then fused scaled_dot_product_attention, reusing existing FFI with no new kernel. MLX fuses the gather into the SDPA read, so gather+SDPA stays within ~15% of contiguous SDPA at <=4k single-sequence context (it climbs with context and batch; the fused-kernel alternative is deferred to #123). Build the gather on the layout-A pool from #118. See docs/adr/0001-paged-attention-gather-vs-fused-kernel.md.

Tasks

• Replace the identity tables with the sequence's real PagedSequenceState block tables in the decode dispatch (src/models/model_owned.rs and the per-model paths in qwen3.rs / llama3.rs / gemma3.rs / llama4.rs).
• Update paged_decode_attention_dense_compat (C++ in cpp/mlx_cxx_bridge.cpp) and the Rust fallback paged_decode_attention_dense_fallback (layers.rs) to gather by physical block id from the pool (Phase 0 option A).
• Apply the same change to the rotating / sliding-window variant paged_decode_attention_rotating_compat.

Acceptance criteria

• Decode output with a fragmented (non-contiguous) block table matches the dense-backend reference within RMS < 5e-3 over 200 steps.
• Supported models (llama3/llama4/qwen3/qwen3.5/gemma3 and their VLM wrappers) are functionally unaffected.

Dependencies

Blocked by Phase 1 (global block-pool tensor storage).

Part of #116