Reject a None loop in Server's constructor instead of crashing later

[afonsojanu]

Fixes #760

uvloop.loop.Server is only ever constructed internally with Server(self), but the constructor accepted anything since Loop loop doesn't say not None. As reported, Server(None) builds fine, and calling close() on it segfaults instead of raising, because self._loop is a typed cdef attribute and Cython's attribute access on it skips the usual None check that a plain Python object reference would get. The crash happens deep inside _unref(), well past where the bad input was actually supplied.

Traced the actual path with the ASan trace from the issue: close() calls _unref() in its finally block, which does self._loop._servers.discard(self). With self._loop set to None, that ends up dereferencing garbage instead of raising AttributeError the way ordinary Python code would, which lines up with the PyType_IsSubtype / __Pyx_PySet_Discard frames in the report.

Marking the parameter Loop loop not None makes the constructor reject bad input immediately with a clean TypeError, which is what the reporter expected in the first place. Checked both call sites in loop.pyx (create_server and the Unix socket path) and they already pass self, a real Loop, so this doesn't affect any real usage.

Reproduced the crash locally first (built from source with the constructor unchanged, confirmed Server(None); server.close() segfaults with exit code 139) before writing the fix, then confirmed the same reproducer raises a clean TypeError afterward. Added test_server_with_none_loop_raises_instead_of_crashing in tests/test_regr1.py, run in a subprocess since without the fix it takes down the whole interpreter rather than failing normally. Reverting just the constructor change makes that test fail with the SIGSEGV assertion, confirming it actually exercises the bug.

Ran the full tests/test_tcp.py suite (112 tests) afterward with no regressions, and flake8 on the touched test file is clean.