sgd-with-adam & weight decay scheduler & Cautious & C4 dataset

[Andron00e]

No description provided.

[Andron00e]

this PR introduces a family of Cautious optimizers and weight decay scheduling (cosine, WSD, stable-decay; with different cooldown shapes).
in addition, this PR include mix of SGD and Adam: Adam trains lm_head, layernorms, embeds by default, all other parameters are optimized by SGD. however this all can be customized further.

TODO:

• customize div_factor of schedulers (default value is 1e2) so we can experiment with down to 10%, down to zero, etc. [done]
• add C4 dataset [ongoing]
• handle the large filename oseerror [ongoing]