Stop __del__ raising on objects whose construction failed

[birkholz-cubert]

What

• Guards all ten __del__ implementations against a handle that was never assigned. Seven of the eight testable handle owners previously raised TypeError out of garbage collection after any failed construction.
• AcquisitionContext.capture(to_internal=True) now reaches the SDK. It passed a Python 0 where SWIG wants an int32_t*, so it raised before the call and the internal queue was unreachable from Python.
• Renames the to_interal parameter to to_internal, with no alias for the old spelling.
• Adds tests/test_handle_lifecycle.py plus three capture tests.

Why

• Every handle owner sets self._handle = None before the call that fills it in, so a constructor that raises leaves a half-built object for the collector. Python prints rather than propagates exceptions from __del__, so the symptom was an unrelated looking traceback appearing after an error the caller had already handled.
• cuvis.h:1886 documents a null result handle as the way to queue a capture internally, so the intent was right and only the marshalling was wrong. SWIG will not coerce a Python int to a null pointer, but it does marshal None.
• The next release is a major version, and correcting a typo is not a feature removal, so the misspelling is simply gone rather than carried through a deprecation cycle.

The queue switch, verified in both directions

cuvis.h:1887 names two consumers of the internal queue, cuvis_acq_cont_get_next_measurement and the worker. Both were checked, with the queue proved empty beforehand so nothing else can be mistaken for the capture under test.

• to_internal=True returns None, and the measurement arrives at the direct consumer and at a Worker with set_acquisition_context. The result is processed to a (275, 290, 51) cube to prove it is a usable measurement and not a handle the SDK never filled in. The queue is empty again afterwards, so exactly one measurement was produced.
• to_internal=False returns an AsyncMesu and puts nothing in the queue. The async result is collected first, so the capture is known to have happened; neither consumer then sees anything. This is the half of the ALL-5828 report that does not reproduce, now pinned by a regression test rather than changed.

Pitfalls

• Breaking for anyone passing to_interal= by keyword; they get TypeError: capture() got an unexpected keyword argument 'to_interal'. Positional callers are unaffected. This wants to land in the major release it assumes.
• The capture tests use the simulated camera fixture, so they are sensitive to queue state left behind by neighbouring tests. Each asserts the queue is empty on entry and drains what it queues.

Verification

• 134 passed against the installed SDK, up from 123.
• The eight new lifecycle tests fail 8/8 on develop and pass 8/8 here.
• Rebased onto develop after Require Python 3.10 and use the annotations it provides #94; no legacy Optional/Union annotations remain, confirmed with ruff check --select UP006,UP007,UP035,UP045.
• ruff check, ruff format --check and scripts/check_changelog.py clean.