gemm: check shape compatibility

[danieldk]

gemm incorrectly allowed multiplication of two matrices that do not have the same inner sizes K:

(N, K) x (K, M)

This leads to reads of uninitialized memory and non-deterministic results.

[danieldk]

Example of undeterministic behavior:

from thinc.api import NumpyOps
import numpy as np

ops = NumpyOps()

a = np.ones((20, 30), dtype="float32")
b = np.zeros((0, 20), dtype="float32")

for i in range(3):
    print(ops.gemm(np.copy(a), np.copy(b))[0][:10])

$ python test.py
[ 7.2064716e+31  1.8474847e+31  1.3043222e-11  6.4460709e-10
  7.3712094e+28  7.2697323e+31 -2.9495326e+12  4.6179864e+24
  7.9981108e+20            nan]
[-1.8212505e+38  7.1954651e+28  9.1228076e+31  1.1784816e+35
  3.2676755e+33  3.4466046e+32  7.8751016e+34 -1.3883587e+33
  1.0641391e+32  5.0848063e+31]
[           nan -3.6532688e+33  1.8077397e+28            nan
  1.3647307e+34  2.5540607e+30            nan -2.0071014e+38
 -1.9679860e+38            nan]

Found while investigating https://github.com/explosion/spaCy/issues/10107 with @adrianeboyd .

[adrianeboyd]

I mainly wondered whether there was a really standard error message used for this, but there doesn't seem to be.