test(point): regression test for Point row/column reference counting (#466)

[blarghmatey]

Summary

Adds a regression test for the Point.row / Point.column reference-counting bug fixed in #466. That PR corrected the getters but did not add a test, and the existing suite can't catch this class of bug (its start_point assertions all use small, immortal-int values).

Background

Before #466, point_get_row / point_get_column returned PyTuple_GetItem(self, i) — a borrowed reference — where a get/set getter must return a new reference. The row/column int was left one refcount too low, so it was freed prematurely and corrupted the pymalloc free list, eventually causing a hard SIGSEGV.

It only reproduces for non-immortal ints — i.e. line/column numbers past CPython's small-int cache (>256) — which is why it surfaced when indexing real, large source files but never with small test fixtures. I ran into it downstream and tracked it to this getter (independently confirmed with AddressSanitizer: a wild TSNode pointer in ts_node_end_byte, one over-decref removed and the crash disappears). See #472 for the original report.

The test

test_point_getters_return_new_references builds a node whose start point is (500, 500) (both values outside the immortal range) and asserts each getter increments the referent's refcount by exactly one. It measures refcounts via tuple subscripting (point[i], which is correctly owned) and asserts before the borrowed reference can do harm, so it fails cleanly on the pre-#466 binding rather than crashing.

Verified:

• Passes on current master (post-fix(point): fix reference counting issues #466).
• Fails (AssertionError: 2 != 3 : Point.row must return a new reference…) when fix(point): fix reference counting issues #466's point.c change is reverted — no segfault.
• Full suite: 61 passed, 20 subtests passed.

The test is skipped on non-CPython implementations (no sys.getrefcount).

[ObserverOfTime]

Thanks but this isn't necessary.