Some code related to train_valid_test_num_samples may be wrong / unused

[odelalleau]

Describe the bug

NB: this was first noticed by @eloialonso

Code related to train_valid_test_num_samples seems weird -- at least for RM training (I haven't checked other algos):

• It is unclear what the initialization to [-1 * cfg.model.global_batch_size] * 3 is meant to do (link)
• It's suspicious that the first element (size of train set) is used when building the validation and test sets (example here)
• It doesn't even seem to be used in the end (in this function)

=> probably that some code clean-up is needed (?)