Fix segmentation fault caused by using libraries depending on pybind11

[jonatanklosko]

This is actually a bug upstream in pybind11, which is a popular library for implementing C extension modules in Python. The bug concerns the GIL, it has been reported a long time ago (pybind/pybind11#2888) and still exists. I found another viable way for us to manage the GIL, such that we don't run into that issue. More details in the code comment.

For the record, below is the code that reliably reproduced the issue for me.

script.exs

IO.gets("PID: #{:os.getpid()} Press enter...")

Mix.install([
  {:pythonx, path: "~/git/pythonx"},
])

Pythonx.uv_init("""
[project]
name = "project"
version = "0.0.0"
requires-python = "==3.13.*"
dependencies = [
  "matplotlib"
]
""")

globals = %{}

# Start a long-running evaluation that releases the GIL, it likely
# runs on the first dirty scheduler thread. This way we can start
# another evaluation and it will run on another dirty scheduler
# thread.
spawn(fn ->
  {_, _globals} =
    Pythonx.eval(
      """
      import time
      time.sleep(20)
      """,
      %{}
    )
end)

Process.sleep(100)

{_, globals} =
  Pythonx.eval(
    """
    import matplotlib
    matplotlib.use("Agg")

    import matplotlib.pyplot as plt
    plt.plot([1, 2], [1, 2])
    """,
    globals
  )

{_, _globals} =
  Pythonx.eval(
    """
    import io
    buffer = io.BytesIO()
    plt.gcf().savefig(buffer, format="png", bbox_inches="tight")
    """,
    globals
  )

(I also removed the interpreter finalization NIF, because we don't use and it's unlikely we ever will, so it's not worth keeping it in sync).

[jonatanklosko]

Looks like it doesn't work on Windows Python 3.10, I will keep investigating.