storing scheduler also

adding separate inference path

simple bug fixes

adding different optimizations as suggested in karpathy GPT-2 video.

updating Bias to False as per Karpathy's

implemented CausalSelfAttention and replaced MultiHeadAttention with it

1. adding functionality of gradient accumulation

significant scale up femto model to get better performance.

implementing multi-head attention in tinygpt

updating README

super minor changes

Move batch sampling into TinyGPT as _get_batch.

minor structural changes

ease of reading

a good version of femto GPT - used andrej's hyperparams