Support fill_value for SimpleImputer with string data

[bodnarbm]

Adds support for setting the fill_value of a SimpleImputer using the constant strategy on string data. This is done by moving the existing guard statement that limits string values of fill_value to be only the value "nan".

Based on some testing, likely this guard statement can be removed completely due to sci-kit learn doing it's own casting checks when fitting data to the SimpleImputer (see https://github.com/scikit-learn/scikit-learn/blob/5600faab452b0adb1f4090219012aa3ae6208c66/sklearn/impute/_base.py#L375-L398). I moved the statement in this PR, in case there was some edge case I missed while doing my own checks and there is some numeric SimpleImputer that allows fill_value=="nan". But let me know if you rather I remove it outright.

Addresses #1122

[bodnarbm]

@xadupre Sorry for the direct mention, but looking at the git history, you might be the best person to ask. Is there anything I can do to help with review on this pull request?

[xadupre]

Sorry for the delay, I was off. Your PR looks good.

[bodnarbm]

No worries.

I will try to clean up the CodeQL reported issues in the PR tomorrow. The CI failures look to be related to other unchanged areas of code, but I am also willing to dive into those too as I have time this week.

[bodnarbm]

@xadupre I went ahead and re-ran the CI jobs on my own fork. It looks like the latest merge from main that I pulled into this branch addressed the remaining test flakiness. I believe the workflows should succeed now for this branch. Thank you for fixing those tests / dependencies in the other PRs you did!