Enable folly

[alexbaden]

This PR is almost identical to #226, but includes one additional commit which propagates the folly compiler definition to the cython code. The previous segfault occurred because cython code was using a std::shared_timed_mutex, whereas the compiled libraries were using folly::mutex.

This does bring up an interesting issue which is the need to propagate compiler definitions to cython more generally. I just did a one-off to get folly working, but I am interested in ways to propagate the set definitions using cmake variables (as a followup).

Also, I found a cython flag to enable gdb. On ubuntu using the conda environment, this gave me enough debug symbols (with regular gdb) to get a stacktrace. I turned the flag off but left it in setup.py.in for future use.

Supersedes #226

[alexbaden]

All issues have been resolved --

1. ENABLE_FOLLY was not being propagated to python code. This resulted in a different mutex type being used in python (std) vs compiled library (folly).
2. The ResultSetTest was using too much memory. This is specifically the large buffers reduction tests, which create a 2^32 size buffer to mimic overflow of a 32-bit group by. This test runs fine on self hosted docker, but does not run on ghrunners, so I disabled it on ghrunners. I was also able to remove the special memory conditions for sanity test (previously smoke test), which do not work for folly because we end up making more use of virtual memory with the Arenas.
3. Modern cmake targets were used in some places, but this resulted in Folly_LIBRARIES being empty in cmake. I use modern cmake targets now and push them back into Folly_LIBRARIES, since other libraries get packed w/ folly (particularly on Windows).

There was one failure in NoCatalogSqlTest but that is a threading issue, and appears unrelated to folly, so this is now ready for review.

I would like to handle the merge as I want to squash everything into one commit and re-word.

[ienkovich]

The previous segfault occurred because cython code was using a std::shared_timed_mutex, whereas the compiled libraries were using folly::mutex.

This is a good catch. We also have other similar conditional declarations, e.g. for threading. We should get all definitions from COMPILE_DEFINITIONS property to be more safe here. I created #297 for that.

[kurapov-peter]

Thanks! There're some minor cmake comments

[kurapov-peter]

I think Folly::folly already exports it, so you don't need it

[kurapov-peter]

I wish we could avoid using any legacy cmake stuff