avoid int32 overflows in indexing

[ev-br]

ILP64 tests will not run on CI. Locally, tested it via $ GULINALG_I32_TEST=1 pytest --pyargs gulinalg on a machine with enough memory.
For a standalone test:

$ cat test_i32_1.py 
import gulinalg as g
import numpy as np

N = 2**31

for i in range(5):
    a = np.ones((1, N+i), dtype=float)
    b = np.ones((N+i, 1), dtype=float)
    c = g.matrix_multiply(a, b)
    print(N+i, c == a @ b)

$ python test_i32_1.py 
2147483648 [[ True]]
2147483649 [[ True]]
2147483650 [[ True]]
2147483651 [[ True]]
2147483652 [[ True]]

[ev-br]

Gave it one more read-through, it still makes sense. Let's land this to allow actual user testing. If any further issues surface in real-life usage, will be happy to fix!