test: add shared gloo process pool for distributed loss tests

[saud5150]

Part of #1982 (PR 1 — draft for early feedback)

Stacked on #1980: the gradient regression test needs that fix (master still has
the raw dist.all_reduce), so #1980's two commits show in this diff until it
merges — I'll rebase onto master then. Review only the four files listed below.

Description

• My change is breaking

Distributed loss behaviour is currently tested either in a separate module
(tests/utils/test_dist__gather__losses.py) disconnected from each loss's own
tests, or by mocking torch.distributed — which only exercises the mock, not the
real collective (exactly the gap that hid #1977 until the mock tests were reverted
in #1980).

This adds a torchmetrics-style session-scoped gloo (CPU) process pool, started
once per session and reused, so a test can submit real multi-rank work cheaply.
Following @liopeer's guidance in #1982:

• tests/ddp_helpers.py (new): USE_PYTEST_POOL flag (default off, folds in the
  gloo/non-Windows checks) and setup_ddp. Adapted from and credited to
  torchmetrics' conftest.py.
• tests/conftest.py: pytest_sessionstart/pytest_sessionfinish start/stop the
  pool; registers the DDP marker. Nothing is enabled in CI.
• tests/loss/test_barlow_twins_loss.py: first consumer — two real 2-rank tests
  (@pytest.mark.DDP + skipif(not USE_PYTEST_POOL)). The gradient test is a
  regression test for [BUG] BarlowTwinsLoss with gather_distributed=True produces gradient which is off by 1/world_size under DDP #1977 and fails on the raw dist.all_reduce.
• pyproject.toml: --import-mode=importlib in addopts. Required because the
  pool uses spawn, whose workers re-import the test module and deadlock under
  pytest's default prepend mode. It is a no-op for the rest of the suite.

Three things worth your review, since you asked to iterate: the submission API
(pytest.pool.starmap(worker, ...), workers return results to the parent to
assert); USE_PYTEST_POOL folding the backend check into the flag rather than a
separate skip condition; and the Barlow tests using identical data per rank
(catches the 1/world_size scaling bug but not heterogeneous-rank bugs).

Tests

• My change is covered by existing tests
• My change needs new tests
• I have added/adapted tests accordingly.
• I have manually tested the change.

Two new tests in tests/loss/test_barlow_twins_loss.py run on the pool; the rest
of the suite is unaffected (flag off → they skip). Verified on macOS/Apple Silicon,
Python 3.9, torch 2.8:

make static-checks                                   # format-check + type-check pass
pytest tests/loss/test_barlow_twins_loss.py          # flag off: 4 passed, 2 skipped
USE_PYTEST_POOL=1 pytest tests/loss/test_barlow_twins_loss.py   # 6 passed (pool runs)

Before/after on the regression test (USE_PYTEST_POOL=1, gradient test only):

raw dist.all_reduce (master):   GRAD MATCH: False   # gradient scaled by 1/world_size
dist_nn.all_reduce  (fixed):    GRAD MATCH: True     # forward is identical either way

Full make test-fast under the new import mode: 4 failed, 1687 passed — the 4 are
the known pre-existing macOS/Python-3.9 env failures (PyAV pickling, a py3.9
assertWarns bug, the existing gloo DCL mismatch), unchanged by this PR and not hit
by CI (Ubuntu 3.7/3.12).

Documentation

• I have added docstrings to all changed/added public functions/methods.
• My change requires a change to the documentation ( .rst files).
• I have updated the documentation accordingly.
• The autodocs update the documentation accordingly.

Test-only change; no public API or .rst docs — no public functions added (the
pool module, hooks, and setup_ddp are internal; the worker helpers use inline
comments per the surrounding test file's style).

Implications / comments / further issues

• Follow-ups within Proper distributed tests #1982 (deliberately out of scope here to keep PR 1 small):
  migrate the existing losses in test_dist__gather__losses.py onto the pool;
  add a heterogeneous-data-per-rank case; enable the pool in CI once it's proven
  green on Linux; revisit pytest.pool (global attr) vs a fixture.
• Not yet confirmed under pytest on Linux CI — the spawn/importlib mechanism is
  platform-independent, but a maintainer sanity-check there would be valuable before
  any CI-enablement PR. Happy to iterate.