perf(inference_models): batch instance-seg RLE mask encoding

[alexnorell]

What

Batch the instance-segmentation RLE mask encoding so post-processing does one device→host transfer per image instead of one per detection.

Why

Instance-seg post-processing encoded masks one detection at a time via align_instance_segmentation_results_to_rle_masks, whose torch_mask_to_coco_rle performs a .cpu() sync per mask (lengths.cpu().tolist() + a GPU scalar read). On Jetson those per-detection syncs serialize the GPU ~2·N times per frame and dominate seg post-processing — a major contributor to the ~50% throughput drop seen running rf-detr-seg on a JetPack 6.2 device vs. detection.

How

• New torch_masks_to_coco_rle_batch(masks) in models/common/rle_utils.py: a single [N,H,W]→[H,W,N] device→host transfer + one vectorized pycocotools.mask.encode.
• Switch the RLE post-process to the already-existing batched align_instance_segmentation_results (the same routine the dense path uses) followed by the batched encode, replacing the per-detection generator loop.
• Applied to: RF-DETR (the non-Triton fallback path; the Triton fast path on main is unchanged) and YOLOv5 / YOLOv7 / YOLOv8 / YOLO26 / YOLACT.

Behavior-preserving

Output is byte-identical — same boxes, same compressed RLE counts. Locked by a new test, tests/unit_tests/models/common/test_rle_batch_encode.py:

• torch_masks_to_coco_rle_batch == per-mask torch_mask_to_coco_rle (random / single / all-zero / all-one / empty).
• align_instance_segmentation_results + batch encode == the per-detection generator, for both boxes and RLE, across no-crop / scaled / static-crop / odd-dims / single / empty cases.

Test plan

• pytest tests/unit_tests/models/common/test_rle_batch_encode.py tests/unit_tests/models/common/test_rle_utils.py → green (CUDA cases skip on CPU).
• No accuracy change expected; this is a scheduling/transfer optimization only.

Draft: opening for review of the approach + scope before marking ready.