Grad accumulation and memory bugfix

[edbeeching]

• Adds command line arg passing to sentiment and toxicity examples
• Adds gradient accumulation as a command line arg #218
• Fixes tensors not being detached from graph when stored in stats, leading to an overuse of memory.
• Updates forward_batch_size -> minibatch size in a number of examples

By the way, I think there may be other places we use excessive memory due to storing attached tensors for too long. I will investigate further.

[edbeeching]

I forgot to run style / quality. I am not on my dev machine at the moment. I will run this in an hour.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.