MatMul LibraryNode expansion failure during codegen

[affifboudaoud]

Describe the bug
Matrix-Matrix multiplication of matrices of the type [N, 1] x [1, M] fails because the MatMul LibraryNode is expanded to a dot product instead of a GEMM. This happens because the subsets are squeezed to decide what to expand the Library Node to.

To Reproduce
On the latest commit in main (01c01cf), run the matrix-matrix multiplication code below:

import numpy as np
import dace as dc

NI = 32
NJ = 36
NK = 1


@dc.program
def mmm(A: dc.float64[NI, NK], B: dc.float64[NK, NJ], C: dc.float64[NI, NJ]):

    C[:] = A @ B


alpha = 0.2
beta = 1.2

A = np.ones(shape=[NI, NK])
B = np.ones(shape=[NK, NJ])
C = np.ones(shape=[NI, NJ])

sdfg = mmm.to_sdfg()

sdfg(A, B, C)

This will output a ValueError: Output of dot product must be a single element

Expected behavior
Output a [32, 36] matrix in the example above.

[tbennun]

Related to #1884 (and will be resolved by it?) @phschaad let’s add this as a test