[BUG] Reference count leak with memory view arguments

[peterbell10]

Describe the bug
When a function has multiple memory view arguments, and the argument call fails because of a type-mismatch then it can leak references to original array being viewed.

To Reproduce
In memleak.pyx:

def foo(double[:] a, double[:] b):
    pass

and in repro.py:

import sys
import numpy as np
from memleak import foo

a = np.random.rand(100)
b = np.random.randint(100)

for i in range(1000):
    try:
        foo(a, b)
    except:
        pass

    if i % 100 == 0:
        print(sys.getrefcount(a), sys.getrefcount(b))

foo is called with a valid array for a but an invalid b. The result is a's refcount is constantly increasing:

4 12
204 12
404 12
604 12
804 12
1004 12
1204 12
1404 12
1604 12
1804 12

Expected behavior
The refcount of a should be constant, like it is for b.

Environment (please complete the following information):

• OS: Linux
• Python version 3.8.8
• Cython version 0.29.23

[da-woods]

I think the problem is that the cleanup:

__PYX_XCLEAR_MEMVIEW(&__pyx_v_a, 1);
__PYX_XCLEAR_MEMVIEW(&__pyx_v_b, 1);

is inside the implementation function foo rather than the wrapper function. Since it fails in the wrapper function and doesn't get to the implementation function then the leak occurs.

It has the signs of "might be fiddly to fix". I think it only applies to memoryviews, but in principle could affect any ref-counted, non-pyobject type added in future