Set env var to hold Keras at Keras 2

[Rocketknight1]

Newer versions of TF use the environment variable TF_USE_LEGACY_KERAS to decide whether to use Keras 2 or Keras 3. We set this immediately in modeling_tf_utils to make sure that no Keras 3 objects sneak in - this should hopefully fix a lot of user scripts that are using tf.keras instead of keras or tf-keras as well.

Fixes #29470

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Why does TF have to be so tricky 🙃

General comment on not overriding environment variables quietly

[amyeroberts]

We shouldn't just override an environment variable like this - this can lead to all sorts of unexpected behaviours on the users end e.g. if they explicitly set it to "0" in their own environment before running a script.

Instead, we should:

• Check if it's already set
• If it's set to 1, then no worries
• If it's set to 0, raise an error saying this isn't compatible with transformers
• If it's unset, set it to 1

[Rocketknight1]

Done! I raise a warning instead of an error because I expect it to be relatively uncommon for people to explicitly set the env var to 0 (as 0 just maintains the default behaviour). Also, it might be possible to use some objects like tokenizers from transformers while still using Keras 3, so I don't want to forcibly raise an error every time modeling_tf_utils is touched.

[amyeroberts]

Thanks for digging into and handling this tricky TF issue!