- Incorporate uv-based CI workflow from codex/migrate-ci-workflow-to-use-uv-projects
- Add [tool.uv] no-binary-package = ['mpi4py'] to force source build
- Fix run-tests.sh to skip pip check when --sdist-mode is passed

Agent-Logs-Url: https://github.com/brainiak/brainiak/sessions/66d28b1c-6cb1-4922-9d90-d7f552a612b6

Co-authored-by: mihaic <165546+mihaic@users.noreply.github.com>

- pydicom: >=3.0.2 (path traversal CVE, was 2.4.4/3.0.1)
- pyOpenSSL: >=26.0.0 (DTLS buffer overflow, was 25.3.0)
- tornado: >=6.5.5 (DoS via multipart, was 6.5.2)
- wheel: >=0.46.2 (path traversal file permissions, was 0.45.1)
- requires-python bumped to >=3.10 (pydicom 3.x requires 3.10+;
  Python 3.9 was already removed from CI matrix in PR #562)
- Regenerate uv.lock with all patched versions

Agent-Logs-Url: https://github.com/brainiak/brainiak/sessions/66d28b1c-6cb1-4922-9d90-d7f552a612b6

Co-authored-by: mihaic <165546+mihaic@users.noreply.github.com>

Continuation of PR #562.

Also fix several CI failures.

## `fastsrm.py`: numpy 2.3 removed `_read_array_header`

`np.lib.format._read_array_header` was a private API removed in numpy
2.3. Replaced with the versioned public API and added a `with` context
manager (the bare `open()` was also causing `PermissionError: [WinError
32]` on Windows when the file stayed open during temp-dir cleanup):

```python
# Before
f = open(path, "rb")
version = np.lib.format.read_magic(f)
shape, fortran_order, dtype = np.lib.format._read_array_header(f, version)
f.close()

# After
with open(path, "rb") as f:
    version = np.lib.format.read_magic(f)
    if version == (1, 0):
        shape, fortran_order, dtype = np.lib.format.read_array_header_1_0(f)
    elif version == (2, 0):
        shape, fortran_order, dtype = np.lib.format.read_array_header_2_0(f)
    ...
```

## CI: MPI platform split — MPICH on Linux

OpenMPI on Ubuntu GitHub Actions runners hangs when `mpiexec` spawns
subprocesses. Split the `mpi4py/setup-mpi` step into OS-specific
variants: `mpich` on Linux, default (OpenMPI) on macOS, `msmpi` on
Windows (unchanged).

## `tests/conftest.py`: `pool_size` fixture logic bug causing EFA abort
+ mpiexec hang

The `pool_size` fixture was supposed to return `1` inside MPI
subprocesses to prevent `fork()`-after-`MPI_Init`. The guard was:

```python
if (multiprocessing.get_start_method() != "fork" and   # always False on Linux
        MPI.COMM_WORLD.Get_attr(MPI.APPNUM) is not None):
    return 1
```

On Linux, `get_start_method()` returns `"fork"` by default, so the
`APPNUM` check was never reached. With MPICH, `pool_size=2` was
returned, `multiprocessing.Pool(2)` forked inside MPI ranks, which
triggered the **libfabric EFA provider** on AWS runners to emit `"Your
job will now abort."` + SIGABRT. `mpiexec` then hung indefinitely →
every affected test hit the 120 s timeout.

Fix: drop the start-method guard so `APPNUM is not None` alone gates the
MPI-subprocess path:

```python
# Whenever inside an mpiexec subprocess, avoid fork() regardless of start method
if MPI.COMM_WORLD.Get_attr(MPI.APPNUM) is not None:
    return 1
```

Also removes the now-unused `import multiprocessing`.

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: mihaic <165546+mihaic@users.noreply.github.com>
Co-authored-by: David Turner <dmturner@princeton.edu>