[Port HIL-SERL] Misc updates for RLPD on real robot

[ChorntonYoel]

What this does

This PR brings changes to a few core files to use the actor/learner training on real robot.
It's not necessarily meant to be merged as a whole. It should at least provide a decent basis for the current work on full robot integration. Feel free to cherry-pick some of the changes, or merge it as a whole when it's ready if good enough.

Changes currently include:
eval_on_robot.py:

• Separation of reward assignment from inference loop so the size of the classifier doesn't impact the inference speed
• Creation of a dataset (to be improved probably), allowing resuming if training is interrupted
• Robot taking back control, so we can intervene, stop the intervention, and intervene again at will during recording.
• Optional manual reset, so we have a few second of teleoperation to reset the robot in the position of our choice
• include an optional image transform to apply on the images before passing them to the classifier

Examples:

Title	Label
Fixes #[issue]	(🐛 Bug)
Adds new dataset	(🗃️ Dataset)
Optimizes something	(⚡️ Performance)

How it was tested

Explain/show how you tested your changes.

Examples:

• Added test_something in tests/test_stuff.py.
• Added new_feature and checked that training converges with policy X on dataset/environment Y.
• Optimized some_function, it now runs X times faster than previously.

How to checkout & try? (for the reviewer)

Provide a simple way for the reviewer to try out your changes.

Examples:

pytest -sx tests/test_stuff.py::test_something

python lerobot/scripts/train.py --some.option=true

SECTION TO REMOVE BEFORE SUBMITTING YOUR PR

Note: Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR. Try to avoid tagging more than 3 people.

Note: Before submitting this PR, please read the contributor guideline.