Segfault with Python 3.7

[tkf]

Running tests with Python 3.7 segfaults. I try to reproduce the issue on Travis in this PR.

The minimal code I found to cause the segfault was:

using PyCall, Dates; convert(PyObject, DateTime(2013))

Added: closes #544

[tkf]

So in all platforms and all Julia versions, we have segfaults.
https://travis-ci.org/JuliaPy/PyCall.jl/builds/420806360
(see, e.g., https://travis-ci.org/JuliaPy/PyCall.jl/jobs/420806366#L881)

[tkf]

Just applying the change in PyDateTime_CAPI struct from https://github.com/python/cpython/pull/5032/files#diff-fbac1e7a0d4bf4957e460fa4560277f4 fixes the segfault.

I didn't touch anything else to add full support of the new C APIs: https://docs.python.org/3/whatsnew/3.7.html#c-api-changes

[tkf]

Actually, TZInfoType wasn't used either and stdlib Dates does not support timezone anyway. So I guess this PR is good to go?

[stevengj]

I'd rather test the default for Conda here.

Maybe change AppVeyor to test Python 3.7 instead of Python 3.4, now that appveyor/ci#2475 made the former available.

[stevengj]

Thanks for looking into this!

[tkf]

Right, I agree that testing against the default Conda.jl setup is important.

I changed Python 3 version in appveyor.yml as you suggested. If #543 CI fails and CI here succeeds, I think it's good to go.

[codecov-io]

Codecov Report

Merging #540 into master will decrease coverage by 0.86%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #540      +/-   ##
==========================================
- Coverage   52.26%   51.39%   -0.87%     
==========================================
  Files          19       19              
  Lines        1502     1502              
==========================================
- Hits          785      772      -13     
- Misses        717      730      +13

Impacted Files	Coverage Δ
src/pydates.jl	62.5% <ø> (ø)	⬆️
src/conversions.jl	49.38% <0%> (-3.4%)	⬇️
src/pyinit.jl	72.72% <0%> (-1.82%)	⬇️
src/numpy.jl	50.3% <0%> (-0.61%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 849adab...60381c2. Read the comment docs.

[tkf]

I rebased this PR onto #544 as it looks like previously tests in x64 AppVeyor were not run properly. The tests were actually run against Python installed via conda since binary incompatibility was automatically detected. Now that I fixed it, Python 3.7 in both x86 and x64 correctly segfault in #544. So this PR should be good to go once the tests here all pass.