fix(vision): capture the functional slice_update return in DeepStack and MRoPE-section loops

[inureyes]

Summary

mlxcel_core::slice_update is functional (wraps mlx::core::slice_update, which returns a new array and never mutates src). Four loops called it as a bare statement and silently lost every write — the same class as the Granite 4 Vision blindness fixed by inject_at_positions in PR #649.

Site	What was dropped	Effect
src/models/qwen3_vl.rs (DeepStack loop)	the whole per-layer visual refinement	quality degradation (main features still enter via inputs_embeds)
src/models/qwen3_vl_moe.rs (DeepStack loop)	same	same
src/models/qwen2_vl.rs (apply_mrope)	the H and W section writes	every head dim kept temporal frequencies; image tokens carried no intra-image spatial position information
src/models/paddleocr_vl.rs (apply_mrope)	same	same

Fix

Capture the return. The DeepStack loops mirror the Granite fix: a contiguous single-image run collapses to one slice_update, disjoint runs fall back to per-row with the return captured. The apply_mrope sites capture the return of each section write.

Validation

• Four co-located regression tests pin the semantics: deepstack_process must add visual embeds at masked positions (qwen3_vl, qwen3_vl_moe); apply_mrope must place per-axis frequencies into their sections (qwen2_vl, paddleocr_vl). All four fail on the pre-fix code.
• Real checkpoints, left-red / right-blue split image through the CLI:
  • qwen2-vl-2b-4bit: "The color on the left side of the image is red, and the color on the right side is blue."
  • qwen3-vl-4b-instruct-4bit: "far left is red ... far right is blue."
  • qwen3-vl-30b-a3b-4bit (MoE): "the color on the left side is red, and the color on the right side is blue."
• paddleocr_vl shares the identical fixed function with qwen2_vl and is covered by its regression test (no local checkpoint on the validation machine).
• Existing qwen3_vl / qwen2_vl / paddleocr unit suites, cargo fmt, and cargo clippy --all-targets -- -D warnings are clean.

Closes #650