How should we handle scalar constants in Elemwise fusions?

[brandonwillard]

We should probably look into the kinds of fusions that turn scalar constants into inputs.

For example,

import aesara
import aesara.tensor as at


x = at.vector("x")
y = at.exp(x / 5)
z = y * 2

f = aesara.function([x], z)

aesara.dprint(f)
# Elemwise{Composite{(i0 * exp((i1 * i2)))}} [id A] 0
#  |TensorConstant{(1,) of 2.0} [id B]
#  |TensorConstant{(1,) of 0.2} [id C]
#  |x [id D]

It's not clear to me that this is any better than—say—making i0 and i1 constants and removing those two inputs.

In other words, why don't we produce the following instead?

Elemwise{Composite{(2.0 * exp((0.2 * i0)))}} [id A] 0
 |x [id B]

Originally posted by @brandonwillard in #1237 (comment)

[brandonwillard]

This issue also affects the applicability of the CAReduce/Elemwise fusion optimization added in #1285, because such graphs (unnecessarily) have multiple inputs, which blocks the optimization—at least until we have the functionality described in #1307.