Fix IntegrityError with unique field and get_or_create

[grigi]

Fixes #140

Had to introduce an extra lock for nested transactions for DB's that support it to prevent network errors with both PostgreSQL & MySQL.

[grigi]

@abondar based this on your original attempt to fix this. Please review.

[coveralls]

Pull Request Test Coverage Report for Build 1338

• 26 of 26 (100.0%) changed or added relevant lines in 4 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage decreased (-0.01%) to 98.062%

---

Totals
Change from base Build 1324:	-0.01%
Covered Lines:	3516
Relevant Lines:	3550

---

💛 - Coveralls

[grigi]

Wow, Py3.6 gets so confused with contexts 😭
I feel we may just have to remove it from CI altogether...
It has been deprecated for a few months now...

Unfortunately it would mean dropping pypy too.

I'll just conditionally skip the test for now.

[grigi]

Another option is to disable all connection pooling on py3.6?

[AEnterprise]

wow didn't expect pooling to be such a major pain on 3.6

but personally i'm more in favor for dropping only specific features on a version like pooling then dropping an entire version (prob best to log a warning on startup though so people are aware)

[grigi]

Py3.6 uses a contextvar backport, and unfortunately it has shortcomings. py3.7+ has extra state in their asyncio C-structs to keep track better.
I'm busy seeing what would be the lowest-impact change for disabling pooling for py3.6. It seems forcing a pool of 1 doesn't do it.
I suspect I would have to use explicit locking like I do for SQLite to get it working right.

[AEnterprise]

i think forcing a shared lock instead of making new locks for concurrent operations should do the trick

that and forcing a poolsize of 1 regardless since it can't use multiple so there is no reason to have any extras open

[grigi]

Well, in the meantime I'm getting hillarious errors, like "cannot rollback; the transaction is not yet started" but investigating further the cause is the dreaded "cannot perform operation: another operation is in progress".

And fixing race conditions fixes that.

[grigi]

@AEnterprise We are discussing the potential deprecation for py3.6 in #303
This PR fixes a bug, and should go in soon.
I'll merge this to master instead of develop, and cherry-pick the windows fix as well and release that. All the new feature work is going in as v0.16

[grigi]

I'll cherry pick the commits instead.