Opt-in FunctionDefinition collection tree node (#3926)

[RonnyPfannschmidt]

Note

Stack: #14855 -> #14856 -> #14769 -> #14857 -> #14858. Review only the last commit.

Adds the collect_function_definition ini option, promoting the internal
FunctionDefinition to a real node in the collection tree.

• hidden (default): unchanged flat layout, no node in the tree
• pedantic: node inserted, function-level markers scoped to it
• messy: node inserted, markers transferred down to each invocation (warns)

Node ids of the individual invocations are unchanged in all modes.

Also introduces nodes.ItemDefinition, the collector-agnostic base for "a
collector standing for one test definition". Concept only here — it carries the
name/selection semantics and nothing else. The generic parametrization protocol
on top of it is #14858.

Node-id selection is fixed as part of this: the definition node carries the bare
name while the parametrization lives on the items below it, so the argument
matcher has to look through the definition. Without this, pytest test_x.py::test_a[1] reports "no match" under pedantic/messy. Covered by a
6 selector x 3 mode regression matrix.

Part of #3926.

[RonnyPfannschmidt]

Heads-up on a defect in this PR as it currently stands: selecting a test by node id is broken in the tree modes.

$ cat pytest.ini
[pytest]
collect_function_definition = pedantic

$ pytest -q "test_x.py::test_a[1]"
ERROR: not found: test_x.py::test_a
(no match in any of [<Module test_x.py>])

Works under hidden, fails under both pedantic and messy. So the PR description's

Invocation node ids stay flat and identical across all modes, so selection, caching and reporting are unaffected.

holds for the node id strings, but not for selection: Session._collect_and_match_nodes walks the tree comparing one name part per level, and the last part is matched as node.name == part + parametrization. With the definition interposed, alpha[1] is compared against the definition's bare name alpha and never matches, so it never descends to the items. -k and full-path invocation are fine — only ::-addressing of a specific test is affected.

The minimal fix is to let the matcher look through an ItemDefinition for the last part (match the bare name, then re-match the same part one level down against the items). I have that plus a regression matrix (6 selector shapes × 3 modes) in #14805, which is stacked on this branch — happy to move it down into this PR instead if you'd rather this one be self-contained.

Worth noting the existing comment right at that spot already anticipates this:

# A non-parameterized arg matches all parametrizations (if any).
# TODO: Remove the hacky split once the collection structure
# contains parametrization.
is_match = node.name.split("[")[0] == matchparts[0]

This PR is exactly what makes the collection structure contain the parametrization, so that split("[") hack is now removable in principle — though doing it properly probably wants the structured NodeId/ParamId from #14758 rather than more string surgery. Will follow up on that.