Gate take_ownership imports on what GraphBLAS can free - #641

A take_ownership=True import handed GraphBLAS any input array with
owndata and writeable set. owndata says an array owns its data, not
which allocator made it: an array from a custom NumPy memory handler (NEP 49)
owns its data too, and GraphBLAS frees what it owns with the allocator it was
initialized with. python-suitesparse-graphblas's letter of 2026-10-04 found
the process aborting (exit 134) when GraphBLAS freed one. Reproduced here
with psg 10.5.1.1, NumPy 2.5.4 and Python 3.15 under PYTHONMALLOC=debug:
an array from a GraphBLAS arena given libc's allocator (so it carries its own
memory handler) raised ValueError from unclaim_buffer after GraphBLAS had
taken it, and the process aborted in Vector.__del__.

Four commits:

Hand GraphBLAS only arrays its allocator can free. Ask psg's
can_unclaim_buffer (new in 10.5.1.1) when it exists, and keep the
owndata check for older psg. An array that fails is copied into memory
from psg's empty, which GraphBLAS allocated, not with NumPy: with
memory_manager="c" (psg initialized before python-graphblas) GraphBLAS
frees with libc's free, so a NumPy copy is no better. Before this, with
psg 10.5.1.1 and memory_manager="c", every ss import raised that
ValueError, take_ownership=False included. The take_ownership
docstrings say what the array needs.

Copy an array given twice to a take_ownership import. GraphBLAS frees
every buffer it takes. The imports copied the values when they were the
same array as one index array, but not other pairs: the same uint64 array
as indptr and col_indices of import_csr was freed twice (exit 134).

Fail a test that frees an owned array with no memory handler. psg's own
conftest guard: NumPy's private _set_numpy_warn_if_no_mem_policy(True)
for the session, so filterwarnings = error fails such a test even where
the allocators happen to agree. It is off on NumPy 1.24 (no setter) and
with psg before 10.5.1.1, whose claim_buffer makes exported arrays with
no handler: forced on with an older psg, 296 tests fail and 13 error,
every one on an array that claim_buffer made. This commit is test
infrastructure only; it can be dropped on its own (its revert applies
cleanly and nothing else depends on it).

Run the one-array-twice import test in a subprocess. A regression there
is a double free, which would abort the whole pytest run (exit 134)
instead of failing one test; the child process imports this tree.

Tests. test_take_ownership_of_foreign_memory imports arrays from a libc
arena (a real foreign memory handler, built without compiling anything) and
checks they were copied; test_take_ownership_asks_suitesparse_graphblas
monkeypatches can_unclaim_buffer to refuse and checks the copy;
test_import_with_libc_allocator (slow) runs imports under
memory_manager="c" with PYTHONMALLOC=debug in a subprocess;
test_take_ownership_of_one_array_twice imports csr, hypercsr and hypercsc
with an index array given twice (in a subprocess). All four fail on the
parent commit with psg 10.5.1.1. The first and third need psg
10.5.1.1 and are skipped with older psg.

Not done: psg's give_buffer context manager. It copies when needed,
commits only if GraphBLAS took the pointer, and unclaims, so it would replace
this and all eleven unclaim_buffer sites. Every import already unclaims
only after check_status, so it adds no safety yet; it restructures all
eleven import paths and needs psg 10.5.1.1 as a floor. Worth it once psg
10.5.1.1 is the minimum, or for the GxB_IS_READONLY load path.