Incorrect broadcasting logic in local_elemwise_alloc

[brandonwillard]

Some broken broadcasting-related logic is causing local_elemwise_alloc to produce graphs that fail to perform simple multiplications:

import numpy as np
import aesara
import aesara.tensor as at
from aesara.compile.mode import get_default_mode


# We're interested in evaluating expressions like the following:
# np.ones((3, 2)) * np.ones((1, 2))

y = at.matrix("y")
x = at.matrix("x")

z = at.ones(y.shape) * x

with aesara.config.change_flags(optimizer_verbose=True):
    z_fn = aesara.function([x, y], z)
# rewriting: rewrite local_shape_to_shape_i replaces Shape.0 of Shape(y) with MakeVector{dtype='int64'}.0 of MakeVector{dtype='int64'}(Shape_i{0}.0, Shape_i{1}.0)
# rewriting: rewrite local_subtensor_make_vector replaces Subtensor{int64}.0 of Subtensor{int64}(MakeVector{dtype='int64'}.0, ScalarConstant{0}) with Shape_i{0}.0 of Shape_i{0}(y)
# rewriting: rewrite local_subtensor_make_vector replaces Subtensor{int64}.0 of Subtensor{int64}(MakeVector{dtype='int64'}.0, ScalarConstant{1}) with Shape_i{1}.0 of Shape_i{1}(y)
# rewriting: rewrite local_elemwise_alloc replaces Elemwise{mul,no_inplace}.0 of Elemwise{mul,no_inplace}(Alloc.0, x) with Elemwise{mul,no_inplace}.0 of Elemwise{mul,no_inplace}(InplaceDimShuffle{x,x}.0, Assert{msg=Aesara Assert failed!}.0)
# rewriting: rewrite constant_folding replaces InplaceDimShuffle{x,x}.0 of InplaceDimShuffle{x,x}(TensorConstant{1.0}) with TensorConstant{(1, 1) of 1.0} of None
# rewriting: rewrite local_mul_specialize replaces Elemwise{mul,no_inplace}.0 of Elemwise{mul,no_inplace}(TensorConstant{(1, 1) of 1.0}, Assert{msg=Aesara Assert failed!}.0) with Assert{msg=Aesara Assert failed!}.0 of Assert{msg=Aesara Assert failed!}(x, Elemwise{eq,no_inplace}.0, Elemwise{eq,no_inplace}.0)

aesara.dprint(z_fn)
# DeepCopyOp [id A] 7
#  |Assert{msg=Aesara Assert failed!} [id B] 6
#    |x [id C]
#    |Elemwise{eq,no_inplace} [id D] 5
#    | |Shape_i{0} [id E] 4
#    | | |y [id F]
#    | |Shape_i{0} [id G] 3
#    |   |x [id C]
#    |Elemwise{eq,no_inplace} [id H] 2
#      |Shape_i{1} [id I] 1
#      | |y [id F]
#      |Shape_i{1} [id J] 0
#        |x [id C]

z_fn(np.ones((1, 2)), np.ones((3, 2)))
# AssertionError: Aesara Assert failed!

mode = get_default_mode().excluding("local_elemwise_alloc")
z_fn_2 = aesara.function([x, y], z, mode=mode)

aesara.dprint(z_fn_2)
# Elemwise{Mul}[(0, 0)] [id A] 3
#  |Alloc [id B] 2
#  | |TensorConstant{1.0} [id C]
#  | |Shape_i{0} [id D] 1
#  | | |y [id E]
#  | |Shape_i{1} [id F] 0
#  |   |y [id E]
#  |x [id G]

z_fn(np.ones((1, 2)), np.ones((3, 2)))
# array([[1., 1.],
#        [1., 1.],
#        [1., 1.]])

I believe this is at least somewhat related to the problematic use of static shape/broadcast information discussed in #1089.