Torchscriptable T5 generation

[joecummings]

Makes GenerationUtils TorchScript-compatible w/ T5 model.

This PR makes the following changes:

• Makes GenerationUtils an nn.Module and TorchScript-compatible
• Modifes the output type of T5 (and variants) and has a simpler update function from encoder and decoder
• Adds parameter with_generation_utils to T5Bundle.get_model() that returns a T5 model wrapped in a TorchScript-compatible GenerationUtils class

Testing:

• Passing all old tests for T5 and GenerationUtils
• Adds a new test for calling get_model with with_generation_utils=True and confirming that it is Torchscriptable and provides the correct results.

[joecummings]

@Nayef211 Thoughts on the with_generation_utils on the model bundle versus having this as a separate class to import?

[Nayef211]

@Nayef211 Thoughts on the with_generation_utils on the model bundle versus having this as a separate class to import?

I think the with_generation_utils is probably a cleaner way to do it and reduces code duplication. I would just make sure to clearly document this parameter so users understand what it does and how it's used.

[joecummings]

@atalman @osalpekar Why is torchdata no longer being picked up in smoke_tests.py?