Feat: added snapshot saving, model downcasting, better dtype/device handling

[Mhdaw]

what is PR does:

• added snapshot saving: it enables continues pre-training by saving optimizer state and hyperparams.
• added model downcasting to lower persicion: it halves the memory footprint, there is performance risk specifically on FP16. if the model is downcasted to BF16/FP16, the logits will be casted to FP32 to avoid numerical instability on loss.
• added better device/dtype handling to automatically select best dtype, data loader params, seed,...
• better torch.compile specification, it checks if its available as well.
• added mini-config for using 135 million params version.
• added automatic lmms-eval installation if true in config
• added requirements.txt for easier installation, added einops, it was missing before!