Bump based CUDA image to ubuntu24.04

[DwarKapex]

Ubuntu24.04 uses python-3.12 as a main interpreter. Unfortunately, not all python packages, we use here, has py-3.12 wheel for amd64/arm64, so need to build the following packages from source:

1. TF-Text
2. Lingvo

Also python-3.12 added a system-wide protection layer (PEP 668) when install packages using pip:

error: externally-managed-environment

× This environment is externally managed
╰─> To install Python packages system-wide, try apt install
    python3-xyz, where xyz is the package you are trying to
    install.

There are at least 2 possible solutions:

1. install everything into venv (the initial solution was proposed by @olupton).
2. System-wide installation by forcing pip installation with env flag PIP_BREAK_SYSTEM_PACKAGES=1

This branch contains both solutions, but collective mind and experience of PyTorch team suggests to finalize the second solution (system-wide installation)

[DwarKapex]

@olupton desperately need your input here =)

[yhtang]

Some builds are still broken 😜

[DwarKapex]

The latest changes includes the following:

1. Installation python packages system-wide (using PIP_BREAK_SYSTEM_PACKAGES=1)
   2.Removing some python packages that installed with apt, but not removable by pip (in pip-finalize.sh)

[yhtang]

Looks mostly good!

Could you please resurrect the .amd64 Dockerfiles, even though we are not using them by default at this time? The concern here is that our lingvo/tftext builds add too many boilerplates to the Dockerfile, and are not easily removable should one day we want to move back to use the upstream wheels. We need a reference here to remind us how simple things could be 😁

[gpupuck]

I don't understand everything you've changed, but I'm good with the changes made for Dockerfile.base and Dockerfile.jax.

[DwarKapex]

Looks mostly good!

Could you please resurrect the .amd64 Dockerfiles, even though we are not using them by default at this time? The concern here is that our lingvo/tftext builds add too many boilerplates to the Dockerfile, and are not easily removable should one day we want to move back to use the upstream wheels. We need a reference here to remind us how simple things could be 😁

Done