Rewrite away reshape that drop dims

[ricardoV94]

Description

This shows up in the following graph:

import pytensor
import pytensor.tensor as pt

x = pt.vector("x", shape=(9,))
out = pt.repeat(x[None], 12, axis=0)
pytensor.function([x], out).dprint(print_type=True)   

# Reshape{2} [id A] <Matrix(float64, shape=(12, 9))> 1
#  ├─ Alloc [id B] <Tensor3(float64, shape=(1, 12, 9))> 0
#  │  ├─ x [id C] <Vector(float64, shape=(9,))>
#  │  ├─ 1 [id D] <Scalar(int64, shape=())>
#  │  ├─ 12 [id E] <Scalar(int64, shape=())>
#  │  └─ 9 [id F] <Scalar(int64, shape=())>
#  └─ [12  9] [id G] <Vector(int64, shape=(2,))>

When a reshape is just adding/dropping dims, we could rewrite as reshape(expand_dims/squeeze(x, ...), shape), which would have led pytensor to remove the reshape altogether.

In general we want reshape to only be used when it actually does something that Dimshuffle does not, as for the eyes of PyTensor it involves a copy (it can't know at compile time, because it would need to know about strides)