[AdnanElAssadi]

The official Qwen3-VL-Embedding and Qwen3-VL-Reranker scripts resize video frames with qwen-vl-utils: each frame is fitted to the patch grid, then the clip to a total_pixels budget at 128–768 tokens per frame. mteb lets the transformers video processor resize instead. It spends one budget on the whole clip, so frames of short clips keep near-native size and small frames are not upscaled. Its cap_pixels_per_frame=True option is closer, but still matches the official size on only 11–31 of the 60 videos below.

This moves UEmbed's tensor resize from #5651 into a shared resize_video and adds the per-frame step it skipped. The Qwen3-VL wrappers go through sentence-transformers, so the resize runs in a small Qwen3VLVideoProcessor subclass before the usual processing, with do_resize=False, the same as UEmbed does.

Patch grid equal to the official one, 60 real videos from 10 MVEB tasks:

Model	main	This PR
Qwen3-VL-Embedding / DNA-VL-STEER	23 / 60	60 / 60
Qwen3-VL-Reranker (its own budget)	24 / 60	60 / 60
UEmbed	38 / 60	60 / 60

Embedding cosine to the official qwen-vl-utils path, 12 videos:

Model	main	This PR
Qwen3-VL-Embedding-2B	min 0.943, mean 0.982	min 0.998, mean 0.999
UEmbed-2B	min 0.967, mean 0.988	min 0.997, mean 0.998

Scores on 32-item slices:

Model	Task	main	This PR
Qwen3-VL-Embedding-2B	MSVDT2VRetrieval	0.956	0.957
Qwen3-VL-Embedding-2B	DiDeMoV2TRetrieval	0.907	0.910
DNA-VL-STEER-2B	DiDeMoV2TRetrieval	0.909	0.919
UEmbed-2B	VATEXT2VRetrieval	0.957	0.961
Qwen3-VL-Reranker-2B	XModBenchT2VReranking	0.531	0.625