Add helper for constructing block matrices

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.

Start by reading this proposal alongside the rewrite work referenced in issue #1044. Define the helper's supported block-matrix shapes and how it should expose zero blocks for chunked matrix multiplication; done should include the helper and the related rewrite behavior.