TF: GPT2 with native embedding layers

[gante]

What does this PR do?

This PR continues the (paused) goal of deprecating our custom TF embedding layers and related code. Previously, we have converted encoder-decoder models (e.g. here), removing TFSharedEmbeddings there and making the necessary adaptations.

In this PR, I make the necessary adaptations for GPT2. The goal is for you, the reviewers, to raise objections in this PR :D All slow tests for TF GPT2 are passing.

Then, the following sequence of PRs will be opened:

1. Remove TFSharedEmbeddings from the other decoder-only models
2. Remove other uses of TFSharedEmbeddings in the codebase (e.g. in tests)
3. Remove resize_token_embeddings and all related functions (it is only used to resize our models' embeddings instantiated with TFSharedEmbeddings)
4. Remove the slow decorator from test_save_load_after_resize_token_embeddings, which will be fixed as a consequence of these changes 🙌

[HuggingFaceDocBuilderDev]

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

[amyeroberts]

Nice - thanks for adding!

[Rocketknight1]

Looks clean to me!

[arivero]

I noticed this change one year later, while doing a demo of LoRA in the embedding layer :-D Nice, but I am always afraid of doing tensorflow operations outside of keras layers. Usually masks are lost, as well as any subtle things that depend of chaining layers. It is not a problem in transformers because the -100 trick is the standard mask and the loss is aware of it.