Speedup the building of Docker images (#18038)

This is a split off #18033 

This uses a few tricks to speed up the building of docker images:

- This switches to use `uv pip install` instead of `pip install`. This
saves a bunch of time, especially when cross-compiling
- I then looked at what packages were not using binary wheels: I
upgraded MarkupSafe to have binaries for py3.12, and got back to Python
3.12 because hiredis didn't have builds for py3.13 with the version we
were using
- The generation of the requirements.txt is arch-agnostic, so I've
switched this one to run on the build architecture, so that both arch
can share it
- The download of runtime depdendencies can be done on the build
architecture through manual `apt-get download` plus `dpkg --extract`
- We were using -slim images, but still installed a bunch of -dev
dependencies. Turns out, all the dev dependencies were already installed
in the non-slim image, which saves a bunch of time as well