fix(testutils): Make _snuba_pool lazy for xdist per-worker URL isolation

[joshuarli]

Problem

In xdist shuffled runs each worker gets its own Snuba container on a different port (1230, 1231, 1232). The module-level _snuba_pool in snuba.py was created eagerly at import time — before pytest_configure ran and set settings.SENTRY_SNUBA to the per-worker URL. All workers therefore shared the default pool pointing at port 1218, causing Connection refused errors under load.

A secondary failure mode: override_settings() restores the original settings.SENTRY_SNUBA value when its context exits, silently reverting any URL override made during configure_for_worker.

Changes

src/sentry/utils/snuba.py

• Replace the eager module-level _snuba_pool (a urllib3 HTTPConnectionPool) with a _SnubaPool lazy proxy class.
• The proxy reads os.environ["_SENTRY_SNUBA_POOL_URL"] (if set) or settings.SENTRY_SNUBA on first use, and rebuilds the inner pool whenever the URL changes.
• Public interface (urlopen) is unchanged; no callers updated.

src/sentry/testutils/pytest/xdist.py

• Rewrite get_snuba_url() to read PYTEST_XDIST_WORKER and SNUBA_PORT_BASE at call time rather than import time, so the correct URL is returned after the worker's env is set up.
• Switch Redis DB allocation from a hard cap of 7 workers to round-robin across 7 slots, so runs with > 7 workers no longer crash.

src/sentry/testutils/pytest/sentry.py

• In configure_for_worker (xdist) and pytest_runtest_setup, write the per-worker Snuba URL into both settings.SENTRY_SNUBA and os.environ["_SENTRY_SNUBA_POOL_URL"]. The env var survives override_settings() resets that would otherwise revert the URL.
• Force a pool rebuild when the URL doesn't match the proxy's cached URL.

Test plan

Verified by running the shuffle-tests-across-shards workflow with 16 shards × 3 xdist workers: Snuba connection errors disappeared after this change.