Replace num_beams with generations_per_sample

[eitanturok]

In composer's build_icl_dataloader function (here), we specify the number of generations per sample with the variable generations_per_sample.

But when we call build_icl_dataloader in llm-foundry here, we specified the number of generations per sample with the variable num_beams. This was confusing because

1. we use different variable names
2. the variable num_beams is also used as a parameter for beam search in inference.

To resolve this, I replaced all instances of num_beams with generations_per_sample in llm-foundry. If generations_per_sample is not specified in the config and num_beams is specified, we will use the value set by num_beams and raise a warning that num_beams is depreciated.

Cheers!

[eitanturok]

Overall, these changes are so that we use the same variable names in both llmfoundry and composer.

Right now, the PR Make CodeEval respect device_eval_batch_size uses the variable generations_per_sample to specify the number of generations per sample and so we change num_beams, which is how we specify the number of generations per sample in llmfoundry, to generations_per_sample.

[josejg]

I believe this PR is redundant with the accompanying foundry PR of the composer you linked: #956

I didn't raise a DeprecationWarning because I do think specifying num_beams was a bug and misnomer, and thus should not be supported any longer as it's a footgun.

[dakinggg]

I'm going to close this PR, it is being covered in #956 . Let me know if you disagree.

[eitanturok]

@josejg -- that makes sense, specifying num_beams is a bug and should not have a DeprecationWarning.

@dakinggg -- no worries, these changes can be implemented in that other PR.