Improve error message for non-broadcasted addition and subtraction

[odow]

Closes #3554

Follow-up to #3557

cc @jd-foster thoughts on this more general error?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (944e5c5) 98.37% compared to head (3518959) 98.38%.
Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #3558   +/-   ##
=======================================
  Coverage   98.37%   98.38%           
=======================================
  Files          37       37           
  Lines        5610     5624   +14     
=======================================
+ Hits         5519     5533   +14     
  Misses         91       91           

Files	Coverage Δ
src/nlp_expr.jl	99.52% <100.00%> (+<0.01%)	⬆️
src/operators.jl	97.07% <100.00%> (+0.15%)	⬆️

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jd-foster]

Looks much more general. I wonder about the related case that also prompted this, #3553, that it might not be obvious that x * LinearAlgebra.I == 1.0 * x. I'm not sure this can be covered here, but might be documented somewhere along the lines of the deleted comment "if you are modifying the diagonal entries of the matrix, do A - x * LinearAlgebra.I(n), where n is the diagonal length."

[odow]

I think in retrospect that supporting UniformScaling was the wrong choice. But yeah, I don't have a good answer for where we should document this. Putting it in an error message that probably happens when the user didn't write I isn't great either.

[blegat]

We can restrict the error message to square matrices. I think it's still nice to include it. It shows why it's ambiguous to add a scalar and a matrix because both would be reasonable interpretations

[odow]

How's this

[blegat]

I like it