[gpt2pre 4] GPT2Preprocessor Layer

[pforderique]

This PR implements the GPT2Preprocessor class with its call method.
Dependencies: #7791 #7806

[UPDATE 7-11-23]
After experimenting with adding generic input/output types to Layer, I noticed the refactoring might be a bit more involved than expected due to a lot of util methods expecting Tensor|Tensor[] or SymbolicTensor|SymbolicTensor[] as their inputs. I have documented this at go/TFJS-layer-generics and will be using the original design outlines below.

[UPDATE 7-10-23]
After discussion with @mattsoulanille today (7-10-23), I will be exploring changes to the base Layer class to allow different input and output types to the call method. This will allow us to support input strings for Tokenizers and any other special i/o needed by the other NLP layers.

[ORIGINAL]
Here is the original plan I had, as implemented in this PR:
I had to deviate from the reference implementation a bit more here because the types can get tricky. Since GPT2Preprocessor extends the Layer class (it really extends Preprocessor, which extends Layer), its overridden call() method must return Tensor|Tensor[].

However, the Keras implementation allows a bit more flexibility by returning the result of pack_x_y_sample_weight(), which returns a packed tuple of x, y, and sample_weight.

Designs Considered:

• I thought about making the TS version packXYSampleWeight() simply return a Tensor of x, y, and sample_weight so that it can be called and directly returned from call() like Keras does, but since x is often used as an object of type {tokenIds: Tensor|Tensor[], paddingMask: Tensor[]}, this cannot be wrapped in a tensor.
• Flattening the results into one large Tensor[] is also an option, but since tokenIds will most like be a Tensor[] itself (since it's the result of calling tokenize() on a tensor of strings), that would mean call would have to support a return type of Tensor|Tensor[]|(Tensor[]|Tensor)[] which starts to get confusing. Not to mention that distinguishing which output is at which position of the tensor would not be super clear.
• Create a new method callAndReturnPaddingMask that just returns a tuple of the tokenIds and paddingMask and have call return the first element (tokenIds) and ignore trying to return y and sample_weight since they aren't mutated by the class. This option is more feasible, however looking ahead, it seems that the GPT2CasualLMPreprocessor class does change y and sample_weight eventually, so it's probably a good idea to figure out how to support this now.
• [DESIGN CHOSEN] I've decided to opt for making a new method callAndPackArgs that returns an array of the arguments packed similarly to the Keras version and call will simply return a subset of these outputs to keep the class layer compatible.

Since GPT2 layers always pass in x as a type of {tokenIds: Tensor|Tensor[], paddingMask: Tensor[]}, I've declared this as a PreprocessorOutput interface and removed Array checking in packXYSampleWeight.

[mattsoulanille]

Looks good. I just have some more tests to add.

[mattsoulanille]

LGTM

[mattsoulanille]

@Linchenn Please take a look when you get a chance. Thanks!

[Linchenn]

LGTM!