[Core][Release] Add Ray TPU image to build pipeline

[ryanaoleary]

Description

It could be useful to build a Ray image with the required JAX and libtpu dependencies to run on TPU (like the ML one defined here) for e2e testing and for users of features like the JaxTrainer. We'd need to add a similar requirements file for TPU to the one defined here for GPU.

Use case

Adding a Ray image with TPU dependencies would be useful for adding e2e tests in the release pipeline for features like the JaxTrainer, or for Ray LLM on TPUs with an image using vllm-tpu.

[ryanaoleary]

cc: @edoakes @liulehui

[edoakes]

@aslonnie is the Ray image & CI czar. Lonnie do you have some pointers on how to do this and/or specific opinions on what the image should and shouldn't contain?

@ryanaoleary if you have an example Dockerfile from an image you've used for testing that would help make it concrete what's needed.

[aslonnie]

is it meant for testing, or we also want to release one?

we already deprecated the ray-ml image.

we have a ray-llm image in the release set that is based on python3.11 and cuda12.8. maybe we can add one ray-llm image for python 3.11 and tpu?

cc @kouroshHakha