lazy dataset loading to save RAM memory

[alisher-ai]

Implemented a lazy data loading inisaac_sim_dataset.py and made it configurable (check base_train_config.gin pretrained_gwm_train_config.gin)

previously;

• the whole dataset was preloaded to the RAM before the training starts,
• during the training, the data was sampled from preloaded data on RAM.

now;

• the file paths are stored on the memory,
• during the training, file paths are read and loaded to the memory by batch.

outcomes;

• saves RAM space,
• faster experimentation (data preloading takes time --> the actual training starts slower --> slower iterations),
• probably a bit slower training, but I did not see any meaningful differences based on single epoch training times,
• lazy_loading can be controlled in config (.gin) files, easy to switch between two modes.

[nv-liuw]

Thanks for the contribution @alisher-ai , LGTM, just a couple of small comments.

[nv-liuw]

LGTM, thanks for the contribution!