Classify plugin startup import failures (#993)

[RonnyPfannschmidt]

Closes #993.

On main a plugin failing to import at startup escapes _main as a plain ImportError: raw traceback, exit 1 — colliding with EXIT_TESTSFAILED by accident — and pytest.main() raises instead of returning its documented exit code. A bare argless ImportError additionally crashes pytest's own e.args[0] handling with IndexError.

#7290 stalled on whether a plugin failing under a conftest's pytest_plugins is a conftest problem or a plugin problem. This PR splits on what actually went wrong instead of on which mechanism asked:

• plugin not found (-p, pytest_plugins, PYTEST_PLUGINS) → misuse → USAGE_ERROR (4), same as a failing conftest.py import
• found but raises while importing — including a missing transitive dependency and a broken pytest11 entry point → plugin defect → INTERNAL_ERROR (3), traceback preserved in the report and on the exception chain (raise ... from) (losing it was the other objection to Issue 993 - improve handling exceptions in conftest #7290)

The split keys on ModuleNotFoundError.name matching the requested spec or a parent package, so a plugin whose own import foo fails is not blamed on the user. conftest.py import failures stay at 4, unchanged.

	before	after
missing plugin via -p / pytest_plugins / PYTEST_PLUGINS	1 + raw traceback	4
plugin raises at import (incl. missing dependency, broken entry point)	1 + raw traceback	3 + formatted traceback
conftest.py fails to import	4	4
pytest.main() in-process	raised ImportError	returns the exit code

Draft because the classification is a judgement call and exit codes are observable API:

1. Is INTERNAL_ERROR (3) right for a broken third-party plugin, or does this warrant a new exit code? wrap_session already uses 3 for plugin exceptions mid-session.
2. Should conftest.py follow the same split? Left alone here to keep the blast radius small.
3. The load_setuptools_entrypoints override widens scope — but without it a broken installed plugin still exits 1 with a raw traceback.
4. Changelog: bugfix or breaking? Filed as breaking per review.

Tests: new TestStartupPluginImportErrors covers every entry path plus both regressions; six existing tests updated from ImportError to UsageError/PluginImportFailure. Full suite and pre-commit (incl. mypy) pass locally.

🤖 Written by Claude Code (Claude Fable 5); reviewed and submitted by @RonnyPfannschmidt.

[RonnyPfannschmidt]

might make sense to wait 1-2 majro releases before putting this in as is

[nicoddemus]

Why wait? Going into the next major release should be fine?

[RonnyPfannschmidt]

Sorry I meant feature release

I recall we have some new deprecation