Switch to slim Python Docker base image for TensorFlow docker builds and tidy up Python Dockerfile

[puneetmatharu]

Note that I currently have

RUN uv pip install -r requirements.txt --no-deps
...
# Install wheel
RUN set -eux && uv pip install "$(basename "$TENSORFLOW_WHEEL")" && rm "$(basename "$TENSORFLOW_WHEEL")"

to mimic what was done before. I'm not sure whether --no-deps should be used on both but the build breaks in that case. It's also the opposite approach to the PyTorch build (--no-deps on the wheel installs and no --no-deps on the requirements installs)... To be discussed...

[jondea]

The --no-deps on the built TS wheels was so that we could have e.g. torchvision~=2.5.0 with torch=~=2.5.0-dev-34er234r23q2342342342323rwrfwfweasgrhgeoiearfy123qb89r7612

[puneetmatharu]

That makes sense. How do we construct our own requirements.txt to go with it? Build the image, install the wheels then do pip freeze? (Just so we know how to keep it up to date)

[jondea]

You could do that, but I think I did it the other way around. I ran each example and installed the minimum required deps to make them work, and pinned just those to minor versions. We only care about the minor versions of things that are script facing, everything below was allowed to change for patch/security reasons. But happy to be persuaded otherwise.

[puneetmatharu]

That's the ideal approach (albeit more costly in time). For now I won't worry about this as it sounds like the current Dockerfile approach isn't invalid and moves us to the simpler Debian build. We can work towards simplifying the requirements.txt as much as possible when/if we do TF binary releases.