Changing num_rollout_samples modifies the validation set in PPO

[odelalleau]

Describe the bug

Technically this may not be a bug, but the current default config setting of model.ppo.num_val_samples (set to ${.num_rollout_samples}) makes it such that modifying num_rollout_samples will also change the validation set, which is bad as it may lead to deriving incorrect conclusions when checking the impact of modifying num_rollout_samples.

Expected behavior

The validation set should remain the same by default when modifying num_rollout_samples.

Additional context

My suggestion would be to deprecate num_val_samples and replace it with limit_val_batches (which is currently missing -- see #27)