bpo-32417: Make timedelta arithmetic respect subclasses

[pganssle]

This fixes both bpo-35364 and bpo-32417. See this comment for a detailed rationale.

By changing timedelta over to returning the subclass it was added to, fromutc now also respects the subclass of the datetime being converted, and consequently so do all the tz=... forms of the alternate constructors.

This will fix many major headaches for people implementing datetime and date subclasses.

https://bugs.python.org/issue32417
https://bugs.python.org/issue35364

https://bugs.python.org/issue32417

[vstinner]

LGTM, but I would prefer to see a review from a second core dev. @serhiy-storchaka, @abalkin, @pablogsal maybe?

[pganssle]

@vstinner Thanks for your comments, sorry for the delay in addressing them - I put it off for a bit because this change likely needs wider discussion and I haven't had time to have that discussion with python-dev before now.

I've addressed your style concerns, rebased against master and started a thread on python-dev for discussion.

[gvanrossum]

What about the C code?

[pganssle]

@gvanrossum That's the first commit, it's here (at the bottom of the diff). It's just a matter of switching to the C function that constructs either a datetime or subclass, depending on the type of self, which is why the diff is so small. This is done in a wrapper function because it's useful in several other places (and also is subject to optimization).

[gvanrossum]

Sorry I missed the C code at first! Reviewing on a phone is not ideal. All looks good. Go ahead and merge!

[bedevere-bot]

@abalkin: Please replace # with GH- in the commit message next time. Thanks!