Add helper for constructing block matrices

[jessegrabowski]

Description

Numpy has np.block, which is gives a nice shorthand for repeated concatenations. For example, say I want to make a block lower-triangle matrix:

$$ D = \begin{bmatrix} A & 0 \\ B & C \end{bmatrix} $$

I can do:

import numpy 
A, B, C = np.ones((3, 4, 4))
B *= 2
C *= 3

zeros = np.zeros((4, 4))
D = np.block([[A, zeros], [B, C]])
print(D)

# Result:
[[1. 1. 1. 1. 0. 0. 0. 0.]
 [1. 1. 1. 1. 0. 0. 0. 0.]
 [1. 1. 1. 1. 0. 0. 0. 0.]
 [1. 1. 1. 1. 0. 0. 0. 0.]
 [2. 2. 2. 2. 3. 3. 3. 3.]
 [2. 2. 2. 2. 3. 3. 3. 3.]
 [2. 2. 2. 2. 3. 3. 3. 3.]
 [2. 2. 2. 2. 3. 3. 3. 3.]]

Obviously you can do this in a million different ways, with np.hstack, np.vstack, np.c_, np.r_r, np.concat, np.stack, etc, etc. But this one is concise and readable.

In the content of pytensor, it's related to rewrites in the same vein as #1044. We could very easily break apart these block matrices and do the matmul in chunks, especially if we see there are zero matrices among the blocks.