Return one frame per requested timestamp in the torchvision_av backend

[rakhimovv]

Thanks for releasing this — a note on one difference between the video backends.

Summary

• get_frames_by_timestamps documents its return as "Frames as numpy array of shape (N, H, W, C)"
  for N requested timestamps. The decord and opencv backends return exactly that; the
  torchvision_av backend seeks to the nearest preceding keyframe and returns every frame it
  decodes up to the last requested timestamp.
• configs/posttrain/libero/libero_all.yaml and configs/posttrain/so101/so101_example.yaml both
  set video_backend: "torchvision_av", so this is the default path for the shipped configs.
• This PR brings the function in line with the copy in
  tutorials/Being-H-EDU/BeingH/utils/video_utils.py, which already carries per-timestamp matching
  and a length check.

Root Cause

The extra frames reach the model because BeingH/dataset/datasets/vla_dataset.py turns each returned
frame into its own ViT image:

for frame in view_group:
    image_tensor = self.vit_transform(Image.fromarray(frame)).unsqueeze(0)
    ...
    packet['num_tokens'] += self.num_image_tokens + 2

With the default obs_indices=[0], the dataset asks for exactly one frame per view per sample, but
the number it gets back — and therefore the number of image tokens packed into the sequence —
depends on how far the sample is from a keyframe.

Measured on two LeRobot datasets, requesting a single timestamp at frame 0, 5, 20, 50 and 100:

dataset (codec)	frames returned before	after
Being-H-EDU_SO101/pick_cube_plate_trimmed (h264)	1, 6, 21, 51, 101	1, 1, 1, 1, 1
IPEC-COMMUNITY/libero_spatial_no_noops_1.0.0_lerobot (av1)	1, 2, 1, 1, 1	1, 1, 1, 1, 1

The size of the effect tracks keyframe spacing in each encode, so it is pronounced on the SO101
videos and mostly invisible on these LIBERO ones.

Test plan

• Single-timestamp requests return exactly one frame on both datasets above (table).
• 10-step LIBERO post-training run on libero_spatial (Being-H05-2B + InternVL3_5-2B +
  Qwen3-0.6B, 1×H100) before and after the change: both complete, checkpoint and
  libero_posttrain_metadata.json written as expected. This is a no-regression check — the
  frame-count table above is what demonstrates the fix.