REGR: using loc to add rows doesn't preserve dtypes

[m-richards]

Potential ideas for solving this. As implemented, _constructor_from_mgr stays nice and clean, but relies too heavily on the pandas internals for my liking (though in an ideal world pandas would expose some way to handle less straightforward dtype concatenation cases).

Edit: This kind of approach is not going to work because of this if else in compat compat:
https://github.com/pandas-dev/pandas/blob/612823e824805b97a2dbe258ba808dc572083d49/pandas/core/dtypes/concat.py#L128-L144
This block assumes to_concat shares a common dtype, and with this trick that's not true

I expect we'll end up using something along the lines of the update to _constructor_from_mgr which is commented out at the moment.

[jorisvandenbossche]

Interesting! That's a nice workaround on our side. I will try to think tomorrow a bit more about how this ideally would be tackled on the pandas side. But on the shorter term, I think this common_dtype trick is an interesting avenue to explore.