Deprecate redundant utilities for extracting constants

[ricardoV94]

Description

In PyTensor we have tho following similar utilities

1. pytensor.get_underlying_scalar_constant
2. pytensor.tensor.basic.get_underlying_scalar_constant_value
3. pytensor.tensor.basic.get_scalar_constant_value
4. pytensor.tensor.basic.extract_constant
5. pytensor.tensor.rewriting.math.get_constant

This PR removes and deprecates all except:

1. pytensor.tensor.basic.get_underlying_scalar_constant_value
2. pytensor.tensor.basic.get_scalar_constant_value

The reason for this distinction, is that the core utility, get_underlying_scalar_constant_value actually works for non-scalar inputs, if it can find a single scalar value underlies a potential n-dimensional tensor (say as in pt.zeros(5, 3, 2)). This is powerful, but can lead to subtle bugs when the caller forgets about it. This was the source of the bug behind #584 and was also likely present in other graphs such as gt(x, [0, 0, 0, 0]) and alike where the repeated condition broadcasts the output of the operation.

The utility get_scalar_constant_value raises if the input is not a scalar (ndim=0) type.

I don't love the underlying distinguishing the two. Perhaps unique would be better.

Both utilities now accept a raise_not_constant which when False (not-default) return the variable as is. I think I would prefer for it to return None

Related Issue

• Closes local_add_neg_to_sub rewrite gives wrong results with negative constants #584
• Related to #