Teleoperation tuning and dataset collection

Goal: Test dataset recording, replay, upload, visualization and teleoperation smoothness

Work Done:
- Tested and tuned full dataset collection pipeline, from teleoperation to online upload and visualization
- Got a lead on lerobot ACT policy observation format
- expects observation.images.{xyz}, however, inside the observation features dictionary, it is not required to use the .images format, lerobot automatically figures out which entries are images
- Collected 10 clean episodes of teloperated test ""
-

Successes:
- test dataset uploaded at https://huggingface.co/datasets/ehalicki/zima-rubiks-cube-test
- full dataset at https://huggingface.co/datasets/ehalicki/zima-rubiks-cube
- can be visualized online at https://huggingface.co/spaces/lerobot/visualize_dataset?path=%2Fehalicki%2Fzima-rubiks-cube%2Fepisode_0

Failures:
- arm kinematics system still has some weird lock ups where it stops responding to new commands because of high error due to large joint state changes
- can increase or remove error threshold?
- can decrease joint error weight?

Next steps:
- Move to switzerland

Time spent: 2 hours