Add ScalarVariable canonicalizations and rewrites

[brandonwillard]

We don't really have any canonicalizations or rewrites for ScalarVariable-based (sub-)graphs.

For example, this would simplify at compile-time in the TensorVariable case:

import aesara
import aesara.tensor as at
from aesara.scalar import int64


a = int64("a")

z = 0 * a

test_fn = aesara.function([a], [z])

aesara.dprint(test_fn)
# mul [id A] 0
#  |ScalarConstant{0} [id B]
#  |a [id C]

Normally, these kinds of graphs wouldn't be generated by a user, so the issue doesn't arise all that often; however, I've noticed some examples of this during debugging that appear to arise from the construction of Composite Ops.

On a related note, we're missing a canonicalization that "compresses" nested Composites:

import aesara
from aesara.scalar import int64, Composite


a = int64("a")

c_op_1 = Composite([a], [a * 2])
c_op_2 = Composite([a], [c_op_1(a) + 3])

z = c_op_2(a)

test_fn = aesara.function([a], [z])

aesara.dprint(test_fn)
# Composite{(Composite{(i0 * 2)}(i0) + 3)} [id A] 0
#  |a [id B]

I'll need to dig up those examples I saw among the Scan tests and demonstrate how these cases arise via our TensorVariable-only interface. In general, we probably want to stick with a TensorVariable-only interface and make as many changes as we can to prevent these situations from arising, but we may also need some of these rewrites regardless.