`Alloc` does not check for runtime broadcasting

[ricardoV94]

Description

import numpy as np
import pytensor.tensor as pt

x = pt.vector("x", shape=(None,))
out = pt.alloc(x, 3, 5)
print(out.eval({x: np.zeros(1)}).shape)  # (3, 5)
grad_out = pt.grad(out.sum(), wrt=x)
print(grad_out.eval({x: np.zeros(1)}).shape)  # (5,)

Interestingly if we do out = pt.alloc(x, 5) then, the rewrites consider that alloc useless and out is replaced by specify_shape(x, 5). AFAICT this was always present in Theano, and was not changed in Aesara.

[ricardoV94]

This was briefly discussed in this old Theano issue: Theano/Theano#2584 (comment)