Fix segfault on repr(OOBucket([... non-Latin-1 thingies ...]))

Fixes #106.

The bug was in not checking the return value of
PyUnicode_AsLatin1String() and passing a NULL to PyOS_snprintf().

The fix is to use PyUnicode_FromFormat() and avoid all the hassle of
manual encoding/sprintf/decoding and even calling repr().

The test fails on Python 3 only, and on non-pure builds only.  I added
PYTHONFAULTHANDLER=1 to the environment to make the failure less
mysteriously silent.

I've also updated tox.ini to use extras=foo instead of deps=.[foo] and
to use factors to set PURE_PYTHON=1 instead of defining an entire
environment.  This now lets you run arbitrary pure Python builds like
tox -e py37-pure.

Drop Python 3.4 support

Stop installing terryfy.

Update tests for persistent 4.4.3

Fix tests against persistent 4.4.

Stop being quite so strict to account for the differences between the
C and Python implementations. But do test that we don't break the new
features when we didn't already have a better form.

Fixes #94

Refs https://github.com/zopefoundation/persistent/issues/92

Avoid spurious "failures" on Appveyor during releases

This way if the sdist hasn't been uploaded yet for some reason, Appveyor
will take care of it.

PowerShell interprets any stderr output from a child process as an
error.  This means warnings from pip about using
not-the-very-latest-version-how-dare-you and warnings from twine about
uploading the *.tar.gz which was probably already uploaded already and
might not be a bitwise identical copy because what even is
reproducibility cause the Appveyor builds to fail, even if the actual
binary wheels we want have been built and uploaded successfully.

*pauses to breathe*

[skip ci]

[skip ci]

python 3.7 binary wheels.

[skip ci]

[skip ci]

* Fix coverage story and add coveralls.io.

* Always be verbose to avoid failing builds because of 10 minutes without output.

* Add badges.

Drop Python 3.3 and add Python 3.7. 

Use operator.index to avoid DeprecationWarning on Python 2

Fixes #79

Slightly faster on CPython:

$ python -m perf timeit -s 'from struct import Struct; s = Struct("i"); unpack=s.unpack; pack=s.pack' 'unpack(pack(123456))'
.....................
Mean +- std dev: 252 ns +- 12 ns
$ python -m perf timeit -s 'from struct import Struct; s = Struct("i"); from operator import index; pack=s.pack' 'pack(index(123456))'
.....................
Mean +- std dev: 235 ns +- 6 ns

Very slightly faster on PyPy:

python -m perf timeit -s 'from struct import Struct; s = Struct("i"); unpack=s.unpack; pack=s.pack' 'unpack(pack(123456))'
.........
Mean +- std dev: 4.24 ns +- 0.12 ns
$ python -m perf timeit -s 'from struct import Struct; s = Struct("i"); from operator import index; pack=s.pack' 'pack(index(123456))'
.........
Mean +- std dev: 4.19 ns +- 0.10 ns

Support PURE_PYTHON at runtime. 

Detect pip installs with 'bdist_wheel' (or manual builds for upload) and don't let the C extensions fail to build unless PURE_PYTHON is set.