Fix several bugs relating to saving offline datasets when not logged in to HuggingFace

[apockill]

What this does

This PR fixes #534 and fixes several bugs I found that all relate to not being logged in to huggingface.

I found a few places that were broken.

1. When --root is set, your dataset was being saved in root/DATA instead of root/hf_username/repo.
2. If you go to control_robot record on an existing dataset, you would get an exception because it was trying to path.mkdir(exist_ok=false), when actually, it should be exist_ok=True.
3. There are still some issues in the training pipeline, but I'm going to leave that for another PR.

How it was tested

1. Record a few episodes
2. Try to continue recording on the existing dataset
3. Train a model (there's still some issues here, I ended up just logging in to huggingface to skip by them. I'll revisit this later)
4. Visualizing the dataset

How to checkout & try? (for the reviewer)

Here's how I've been running control_robot record

record
--robot-path lerobot/configs/robot/myarm.yaml
--root "data"
--repo-id username/dataset_name
--single-task "Cool task description, wow!"
--fps 20
--push-to-hub 0
--resume 1
--local-files-only 1

[Cadene]

@apockill We changed the behavior of root to be more explicit.
You would need to provide --root "data/username/dataset_name".
Could you try with that?
cc @aliberts

[apockill]

Ahh got it. I'll try that soon and report back. Thanks!