Matmul test coverage

[newling]

As discussed in the sync, we will start with something like

dimset = [8, 16, 132, 64, 128, 256, 512, 1024, 2048]
seed = 42
n_tests = 10
for in range(n_tests):
  m = np.random.sample(dimset)
  n =  np.random.sample(dimset)
  k = np.random.sample(dimset)
  run_matmul_test(m, n, k)

We will choose n_tests as large as possible (i.e. setting it 1 larger results in a failing test) and create a task to resolve the issue encountered with n_tests + 1. When that issue is resolved, we'll increase n_tests to the largest possible value.

We expect n_tests to be low initially, as we know our coverage is not good.

When n_tests gets sufficiently large we can freeze it to not make CI too slow. Also at this point we can consider a nightly job, and also consider unfreezing seed.