Add specialization for subtraction and addition with negative terms

[ricardoV94]

x - (-y) -> x + y
x + (-y) -> x - y 
-x + y -> y - x

Adding these during canonicalization conflicted with a couple of other rewrites.

1. local_neg_to_mul tries to convert -x -> -1. * x, so that the local_mul_canonizer can do its thing.
2. When it comes to constants, the local_add_canonizer will try to place them on the left of the graph for "constant folding" and leave them as an add, even if they are negative. For example x - 5 -> -5 + x.

The new rewrites were registered or designed not to conflict with these preexisting rewrites

Closes #546

[codecov]

Codecov Report

Merging #549 (1846741) into main (77df667) will increase coverage by 0.00%.
The diff coverage is 96.00%.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #549   +/-   ##
=======================================
  Coverage   74.09%   74.10%           
=======================================
  Files         174      174           
  Lines       48605    48622   +17     
  Branches    10340    10348    +8     
=======================================
+ Hits        36014    36030   +16     
  Misses      10305    10305           
- Partials     2286     2287    +1     

Impacted Files	Coverage Δ
aesara/tensor/rewriting/math.py	86.34% <96.00%> (+0.08%)	⬆️

[brandonwillard]

This looks really good! I have some comments/questions about the special DimShuffle logic, but that's it.

[ricardoV94]

I should take a look if somethings can be removed from here after this PR:

aesara/aesara/tensor/math_opt.py

Line 2686 in 391e2cf

local_one_plus_erf = PatternSub(

There seems to be a lot of redundancy going on

[brandonwillard]

Looks like CI was stuck, so I rebased it.

[rlouf]

Rebased the branch on main and resolved the merge conflicts as this looks still relevant.