Broken GEMM C implementation when using NumPy >= 1.23

[brandonwillard]

It looks like numpy/numpy#21477 (i.e. NumPy >= 1.23) is causing issues in the C implementation of GEMM.

Here's some example test output illustrating the difference:

# Using NumPy 1.23.3
tests/tensor/test_blas.py::TestGemm::test_shape_0
x.shape=(0, 5), x.strides=(0, 0)
y.shape=(5, 4), y.strides=(16, 4)
z.shape=(0, 4), z.strides=(0, 0)
...

# Using NumPy 1.22.3
tests/tensor/test_blas.py::TestGemm::test_shape_0
x.shape=(0, 5), x.strides=(20, 4)
y.shape=(5, 4), y.strides=(16, 4)
z.shape=(0, 4), z.strides=(16, 4)