fix(docker): install python3.12-dev so Triton can JIT in the container (#453)

[Anai-Guo]

Is your pull request related to a problem? Please describe.

Fixes #453. Starting the container logs a burst of build failures before the server comes up:

/tmp/tmpe68t6sye/cuda_utils.c:6:10: fatal error: Python.h: No such file or directory
    6 | #include <Python.h>
      |          ^~~~~~~~~~
compilation terminated.
/opt/venv/lib/python3.12/site-packages/fla/utils/_device.py:100: UserWarning: Triton is not supported on current platform, roll back to CPU.

triton is a declared dependency of both the cu12 and cu13 groups
(pyproject.toml L74-76 / L101-102), so it is always in the image. Triton compiles its
cuda_utils extension at runtime, which needs the CPython development headers.

Both Dockerfiles install python3.12, python3-pip and python3.12-venv, but not
python3.12-dev — so Python.h is not in the image. build-essential is already there,
so the headers are the only missing piece.

Putting them in the venv would not help: sysconfig reports the interpreter's include
path from the base prefix even inside a venv, which is where Triton looks —

$ /opt/venv/bin/python -c "import sysconfig; print(sysconfig.get_paths()['include'])"
/usr/include/python3.12

and on Ubuntu 24.04 that path is owned by python3.12-dev (via libpython3.12-dev):

$ dpkg -S /usr/include/python3.12/Python.h
libpython3.12-dev:amd64: /usr/include/python3.12/Python.h

Why should this feature be added?

The failure is not only log noise. The compile error is what flash-linear-attention's
Triton probe hits, and it responds by rolling back to CPU — so an image that ships Triton
on purpose silently runs without it.

Both docker/Dockerfile and docker/Dockerfile.cu13 have the identical apt-get block
and the identical gap, so both are fixed here.

Examples

Before:

$ docker compose -f docker/docker-compose.yml up
tabbyapi-1  | /tmp/tmpe68t6sye/cuda_utils.c:6:10: fatal error: Python.h: No such file or directory
tabbyapi-1  | compilation terminated.
tabbyapi-1  | .../fla/utils/_device.py:100: UserWarning: Triton is not supported on current platform, roll back to CPU.

After: Python.h resolves, the JIT build succeeds, and the rollback warning is gone.

Additional context

One package added per Dockerfile; no change to the venv, the install steps, or the
entrypoint. The added layer is a few MB. The existing Docker smoke test workflow builds
both files on any docker/** change, so this PR exercises both images.

I do not have a CUDA host to run the image on, so I verified the mechanism rather than the
container: the sysconfig include path and the dpkg -S ownership above were both checked
on Ubuntu 24.04 with CPython 3.12. Worth a maintainer running the image once to confirm the
warning is gone end to end.

🤖 Generated with Claude Code