Disable TLS for MINGW due to issues with the order of freeing up memory.

[darinpp]

There seems to be an issue with the way MINGW frees up TLS memory. More specifically - the TLS is freed up before the destructors on the objects residing in TLS are called. In our case - we were observing a destructor accessing a block of memory that was already freed.

Here is a report from end uses that seems to be related https://sourceforge.net/p/mingw-w64/bugs/727/

Fixes cockroachdb/cockroach#40936
Fixes cockroachdb/cockroach#40925

---

This change is 

[petermattis]

though I'm curious why we're trying to run destructors at process exit vs a more abrupt termination.

Reviewable status: 0 of 1 files reviewed, all discussions resolved (waiting on @ajkr and @petermattis)

[darinpp]

I doubt that abrupt exit is the best things to do. I could imagine trying to log some information at termination time. One place where I saw the issue is here:

rocksdb/monitoring/perf_context.cc

Line 36 in afdceb8

ClearPerLevelPerfContext();

though I'm curious why we're trying to run destructors at process exit vs a more abrupt termination.

Reviewable status: 0 of 1 files reviewed, all discussions resolved (waiting on @ajkr and @petermattis)

[petermattis]

I doubt that abrupt exit is the best things to do.

It's pretty common technique to avoid trying to cleanly tear down some complicated setup. On the other hand, it might paint us into a corner with regards to doing real CI tests on Windows in the future.