Do not report a registered type as "Unknown type" in get_cuda_native_handle

[LeSingh1]

get_cuda_native_handle() wraps both the registry lookup and the getter call in a
single try:

    obj_type = type(obj)
    try:
        return _handle_getters[obj_type](obj)
    except KeyError:
        raise TypeError("Unknown type: " + str(obj_type)) from None

The except KeyError is there for one thing: "this type has no registered getter". But
_handle_getters[obj_type](obj) is two operations, so a KeyError raised inside the
getter is caught too — and then the diagnosis is wrong twice over:

• the reported type is registered, so "Unknown type" sends the reader looking in
  exactly the wrong place;
• from None sets __suppress_context__, so the traceback that would have shown the real
  failure is not printed at all.

Reproduced (no GPU needed):

>>> _add_cuda_native_handle_getter(Registered, getter_that_raises_keyerror)
>>> get_cuda_native_handle(Registered())
TypeError: Unknown type: <class 'Registered'>

Fix

Move the getter call out of the try, so only the dict lookup is guarded. The
unregistered-type path is byte-for-byte unchanged, and the existing
test_get_handle_error still covers it.

This is a latent mis-diagnosis rather than a crash — today's registered getters are
generated one-liners — but it is the kind of "error path that fires for the wrong reason
and then destroys the evidence" that #2122 is about, and the fix is three lines with no
behaviour change on either existing path.

Tests

One test added next to the existing test_get_handle / test_get_handle_error pair in
cuda_bindings/tests/test_utils.py: it registers a getter that raises KeyError (via
monkeypatch.setitem, so the global registry is restored) and asserts the KeyError
propagates. Fails on main with TypeError: Unknown type: <class 'Registered'>.

Verified against upstream/main and with the change; the happy path and the
unregistered-type path both keep passing either way. ruff check and
ruff format --check clean.

Note: #2541 also touches cuda_bindings/tests/test_utils.py, appending at the end of the
file; this change lands mid-file next to the existing handle tests, so the two are
independent. Happy to rebase whichever lands second.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.