perf(multimodal): optimize Qwen vision runtime

[yechank-nvidia]

Summary

• Reduce Qwen vision CUDA graph replay overhead by reusing prepared grid metadata and updating only active buffer regions.
• Avoid redundant tensor concatenation when item tensors already share contiguous storage.
• Reduce encoder output cloning and preserve views when packed outputs are already in request order.
• Coalesce adjacent vision scatter ranges and avoid unnecessary deepstack allocation/copies when deepstack is disabled.
• Preserve the existing encoder output contract and chunked-prefill behavior.

Test Plan

• Verified CUDA graph batch selection, metadata padding, replay buffer updates, output ordering, and cloned-output ownership.
• Verified image/video embedding assembly and chunked-prefill deepstack behavior.
• Compared optimized Qwen outputs against the existing path for exact equality.
• Ran Qwen3.5-4B image and 720p24 10-second video E2E at concurrency 1, 8, and 32.
• Confirmed E2E performance remained within run variance; measured embedding assembly decreased from approximately 0.21-0.29 ms to 0.08-0.13 ms.
• pytest -q test/runtime/test_encoder_cudagraph_grid_rows.py test/runtime/test_qwen_vision_runtime_optimizations.py
• pre-commit run --all-files