Git-t5

* Remove extra config argument from model loading function

* Move code-specific tools into `git_t5.utils.code_tools` submodule
* Add `encode_input` function to test out a model inference in runtime

* Fix total number of batches calculation
* Fix `datasets` package version lock

* Add more BPE tokenizer configuration options
* Avoid default NMT normalization

* Hydra-based training configuration
* T5 model pretraining script
* T5 tokenizer pretraining script
* Code pre-tokenization script