Fixing entropy calculation

[abaheti95]

I made a mistake when logging entropy by only looking at the actions rather than the entire distribution.
This fixes that mistake and now does correct entropy calculation

mlflow of old and new entropy logging

Update

I've modified the PR to include more functionalities:

1. Getting the sequence and token entropy of the sequences
2. seq entropy becomes the loss term
3. Token entropy is used to get different percentiles of token entropies throughout the batch

[dakinggg]

Since fixing a bug in a narrow, fairly self contained piece of the codebase, lets add unit tests while we're here?

[gupta-abhay]

Since fixing a bug in a narrow, fairly self contained piece of the codebase, lets add unit tests while we're here?

+1 to this

[abaheti95]

Okay... modified the code to calculate both token and sequence entropies and added unit tests for both.
Also added piping for entropy loss from @jdchang1 's entropy loss branch

[jdchang1]

LGTM! Thanks for the fix and tests

[gupta-abhay]

mostly lg, some nits!