Closes #3579.

Restore the previous PyThreadState_Get() calls in the refnanny, but with a comment explaining the purpose of the "useless" calls.

Turn all legacy struct fields in the NumPy ndarray type into properties and document them (copied from NumPy docs).
Adapt the 'nogil' declarations according to the ones in NumPy and add some missing functions.
See https://github.com/numpy/numpy/pull/16170

* Support for versioned pxd files like "lib.cython-30.pxd" for a Cython 3.0+ version.
See https://github.com/cython/cython/issues/3573

* Fix test in Py2 by avoiding namespace packages in favour of a normal package.

Remove "PYTHONHOME" setting from embed-test since it seems to break the I/O encoding setup in (at least) Py3.8+ venvs.

See https://github.com/numpy/numpy/pull/16170

Add more 'nogil' declarations to NumPy functions that do not require the GIL (according to the declarations shipped by NumPy itself).
See https://github.com/numpy/numpy/pull/16170

Remove the outdated getbuffer/releasebuffer implementations for the NumPy 'ndarray' since there are probably no NumPy installations out there anymore that do not support the buffer protocol themselves and are still worth supporting.

See https://github.com/numpy/numpy/pull/12284#discussion_r420378155
See GH-3573

Reverse the tuple item order in the utility code caching key to move the most selective item first (just in case if we ever get dict collisions).

See https://github.com/cython/cython/issues/3573

Tighten "numpy_attributes" test to assert that it's actually using a property call and not an attribute access.
Also add a test for the newly added "ndarray.data" property.

Mark the embedded C main() function in the embedding test as "extern C" in C++ to see if that fixes the C++ builds in MSVC.

Fix a C compiler warning about unused cleanup code in buffer using functions that always raise (found by clang in the bufaccess.pyx test).

* Rewrite C property support (GH-2640) based on inline C methods.
Supersedes GH-2640 and GH-3095.
Closes GH-3521.

* Test fix for `numpy_parallel.pyx`: avoid depending on whether "nd.shape" requires the GIL or not.

* Turn NumPy's "ndarray.data" into a property to avoid direct struct access.

* Make "ndarray.size" accessible without the GIL.

Prevent VerboseCodeWriter from inserting nul-bytes into the C file by switching to a more "standard" way of inserting code at a later time.