Fix online training

[Cadene]

What does this PR do?

• Use online_steps=2 in end-to-end test, since 1 wasn't enough to capture the concatenation logic between first episode and second episode
• Use env.episode_length=2 in end-to-end test to make it faster
• Match the features type between the output of eval_policy and the offline/online datasets

Minimal code example to checkout and try

python lerobot/scripts/train.py \
policy=tdmpc \
env=xarm \
wandb.enable=False \
offline_steps=1 \
online_steps=2 \
env.episode_length=2 \
eval_episodes=1 \
device=cpu \
save_model=true \
save_freq=2 \
policy.batch_size=2 \
hydra.run.dir=tests/outputs/tdmpc

How did you tested the changes?

CI + breakpoint

[alexander-soare]

One comment, otherwise LGTM. Note: for this review I am happy with a brief skim and knowing that it's tested and works for me. I know we are refactoring eval.py soon.