Pin the selenium BiDi surface with a guarded test

[vishnuv688]

Part of #277

The adapter reaches past the public API into driver.network.conn,
bidi.network.NetworkEvent and bidi.session.Session. tests/test_bidi.py runs
entirely on fakes, so a selenium upgrade that moves any of them passes CI green
and degrades at runtime. Selenium PR 17761 regenerates that layer from a shared
schema, which makes the move likely rather than hypothetical.

Acceptance criteria

A test, skipped when selenium is absent, asserts those
three imports and attributes resolve on the installed selenium. It runs in
python.yml and fails loudly on a selenium bump that moves them.

Size: S.

[vishnuv688]

Partly addressed, and the CI half is provably unmet

#317 added one guarded test of this shape — test_pin_registers_without_a_browsing_context — because the preload depends on driver.script.pin() registering globally. pin()'s docstring says "Pins a script to the current browsing context", but it forwards no contexts argument and BiDi reads that as every context. That reliance is undocumented, so a test asserts it: if selenium ever scopes pin to one context, the test fails rather than the preload silently covering only the first document.

That does not close this issue, on either axis:

1. The named internals are still unpinned. This issue is about driver.network.conn, bidi.network.NetworkEvent and bidi.session.Session. None of those has a guard; tests/test_bidi.py still runs entirely on fakes.

2. The acceptance criterion "it runs in python.yml" cannot currently be met. The job installs nothing and runs PYTHONPATH=src python -m unittest discover -s tests, so selenium is absent and any test of the real surface must skip — which is exactly what the new one does:

Ran 230 tests in 3.812s
OK (skipped=1)

So the guard exists but never runs where it is supposed to protect. A selenium bump that moves these attributes still passes CI green, which is the failure this issue was filed to prevent.

The fix is therefore partly a CI change, not only a test: the Python job needs the adapter's own runtime dependency installed (pip install -e . or equivalent) before a guard of this kind is worth anything. That also makes the existing pin() guard live, and would let several other tests exercise real selenium rather than fakes.