fineweb10B/gpt2 recipe, and supporting changes

[francoishernandez]

This is blatantly inspired from karpathy/llm.c#481.

Some changes in the process:

• support adamw optimizer;
• support cosine style decay (linear warmup + cosine decrease to 0);
• removed param_init_glorot in favor of a more generic param_init_method;
• check all recipes config in tests.

TODO:

• proper README
• full reproduction run / multi-gpu run + logs/charts
• hellaswag benchmark (Add gpt2 converter, hellaswag eval tool, misc fixes #38)