Allow flaky conditions to inspect failure exceptions

[ryux1]

Closes #230

This lets flaky conditions make one rerun decision for an entire test attempt from the exceptions that caused its failed phases:

• one-argument callable conditions receive each failure exception;
• existing zero-argument callable conditions remain supported;
• string conditions inspect the exception through the reserved error name;
• setup, call, teardown, and subtest failures retain their corresponding exceptions;
• an attempt is rerun when any failed phase matches, with each failure evaluated at most once.

ExceptionInfo stays on the worker-local test item, so xdist never serializes it, and the cached exception state is cleared after every attempt. Condition evaluation errors become non-fatal warnings even when warning filters promote warnings to errors.

Validation:

• pytest 8.2, 8.3, and 8.4: 206 passed, 19 skipped in each environment
• pytest 9.0 and 9.1: 216 passed, 9 skipped in each environment
• pytest 9.1 with pytest-xdist: 225 passed
• pre-commit run --all-files passed

[icemac]

The test failures against pytest-main are also on master and need to be fixed there before we can make progress here.

[icemac]

master is fixed now, could you please resolve the conflicts?

[icemac]

Thanks for working on #230 — the xdist-safe approach of keeping ExceptionInfo on the worker-side item is the right idea. A few things need fixing before this can go in; two of them abort the whole run with INTERNALERROR, and one makes the feature a no-op for failed subtests. Details inline.

— Comment created by Claude