Sorry, I think we shouldn't be changing `ast.py` this way. To reiterate my review comment from the PR:

I think we should just let this go and instead advertise that if you put new synthetic nodes in a tree, you have to either:
- provide `lineno` and `col_offset` yourself; or
- use `fix_missing_locations()` to fill out the information for you.

The problem with the logic that I commented on twice now (on the PR) is one reason for my opinion. Another is bigger though:

Suppose user code is replacing nodes or just inserting new ones. This operation not only creates nodes without location attributes. It also in all likelihood *moves* locations of type comments and those aren't updated (for example, `_type_ignores[123]` should now really be `_type_ignores[125]` to still point at the same tokens). So user code shouldn't be naive when it comes to missing locations, and we definitely shouldn't paper over missing locations.