rewrite test helper

[ricardoV94]

This PR adds a helper to streamline our rewrite tests, according to what I these days feel is the best strategy:

1. Write a graph
2. Rewrite
3. Check the rewrite matches what we expect
4. Eval both original and rewritten without any extra compilation

For instance, the test in #2101 would look like:

@pytest.mark.parametrize("b_ndim", [1, 2], ids=lambda x: f"b_ndim={x}")
def test_solve_of_inv_to_matmul(b_ndim):
    X = pt.dmatrix("X")
    b = pt.dvector("b") if b_ndim == 1 else pt.dmatrix("b")

    out = solve(pt.linalg.inv(X), b, b_ndim=b_ndim)

    result = utt.rewrite_test([X, b], [out])
    result.assert_equivalent_computations([pt.dot(X, b)])

    rng = np.random.default_rng(42)
    X_val = rng.random((4, 4)) + np.eye(4) * 4
    b_val = rng.random((4,) if b_ndim == 1 else (4, 3))
    result.assert_numerical_close([X_val, b_val], rtol=1e-6)

I really don't like op counts, because it can miss stuff like AdvancedSubtensor become AdvancedSubtensor1, or Blockwise(Dot) become Dot, or Gemm, and it seems like we are optimizing stuff when we are not. Still I added the helpers to count ops...

The main goal though is to reduce friction / have a baseline for contributions

[jessegrabowski]

Why do you think a numerical test should be standard after a structural comparison? It seems like the structural should be sufficient in most cases.

[ricardoV94]

Why do you think a numerical test should be standard after a structural comparison? It seems like the structural should be sufficient in most cases.

I also like structural check but we've been bit by that in the past. We got the structure we expected but there was a flaw in it. The numbers are less likely to be fooled

[ricardoV94]

Example the square(sqrt) rewrite we got the wrong order recently. Needed nan in the test but you get the idea