[WIP] fix from_type under Python 3.7

[Zac-HD]

Closes #1264, closes #1148.

• Fixes two new deprecation errors in our tests (related to asyncio and collections.abc), and a test failure related to the new UTF-8 mode (by accepting the escape code or the character itself)
• Cleans up after the from_type puppy again; given that the typing module is still provisional I expect to do this again for 3.8 😭. This means adding TypeVar support because bare generics now have explicit typevars as their contents at runtime (ugh), and adding yet another set of compatibility shims and conditional logic. It's not pretty but it does all work.
• dataclasses already work via the existing annotation-based machinery - it's a miracle of unusually restrained design, but there you are.

[DRMacIver]

Whoops. sorry, I fucked that up! Wrong window open when I clicked the close and comment button.

[Zac-HD]

CC @DRMacIver for review.

[DRMacIver]

I'm feeling a bit 🤷‍♂️ on how to review code on this feature TBH. The workarounds required to make it work seems to be getting increasingly involved and hacky, and I'm not very happy about that, but I'm not really sure what to suggest doing about it (other than maybe dropping the feature until the typing API stabilises, which would be sad).

Anyway, tests pass, which is an improvement on what we had before, so I'm going to rubber stamp it. I've commented on two things that I'd like addressed (one might be just me misunderstanding the code) but I don't need to re-review afterwards.

[DRMacIver]

I feel like we should maybe try to contain our understandable exasperation with the CPython core team in our release notes.

[DRMacIver]

It doesn't look like typing is guaranteed to be in scope here? The above import might have failed and been caught.

[Zac-HD]

The workarounds required to make this feature work seem to be getting increasingly involved and hacky, and I'm not very happy about that, but I'm not really sure what to suggest doing about it (other than maybe dropping the feature until the typing API stabilises, which would be sad).

I'm also pretty unhappy about it, but agree that there's not much to be done. Basically there's just a lot of churn in the runtime behavior of the typing module, so we end up with a lot of fiddly special cases for odd objects under specific versions. Dropping support for from_type(some_generic_type) under 3.5 and 3.6 would probably each buy about a 25% reduction in the total amount of code - I'd rather not do that when it's all working, but it wouldn't be hard to talk me into dropping 3.5 as it barely works anyway.

I'll merge this as an improvement on the status quo, but agree that this isn't great either 😕