travis: update PyPy jobs to use more recent versions

tests: fix `test_pip_upgrade_from_source` on Python 3.4

Do not test pip's master on 3.4, as support for it has been dropped.

Perform a more thorough check of package_data structure

Ensure the dictionary values are lists/tuples of strings.

Fix #1459.

value=None raises TypeError

DistutilsSetupError: 2 must be a list of strings (got None)

Include the file path in get_metadata() UnicodeDecodeErrors.

Fix CI

Use license classifiers rather than the license field.

stop using deprecated HTMLParser.unescape

HTMLParser.unescape is accessed even when unused - this will cause an
exception when `HTMLParser.unescape` is removed in Python 3.9.

The license field has a 'free' format, while the classifiers are unique
identifiers, similar to SPDX identifiers. In the documentation, we
should therefore showcase the use of classifiers.

tests: unpin pytest

The new releases for pytest-fixture-config and pytest-shutil are
compatible with pytest>=4.0.

Force metadata-version = 1.2 when project urls are present.

Closes: #1756

Migrate constants from setup.py to setup.cfg

This also makes wheel an unconditional setup_requires dependency.

Closes: #1697

Fix CIs.

Fix the `list_dependencies_command` so arguably the most important
packages are not hidden (pip, setuptools, and wheel).

Update virtualenv so the tox environment use a more recent version
of pip (and probably setuptools/wheel too).

* ignore distutils' warning (`Unknown distribution option: 'python_requires'`)
* fix test on Windows

build_meta: fix 2 issues with `build_wheel` / `build_sdist`

Fix the following cases:

* `build_sdist` is called with another sdist already present
  in the destination directory

* `build_wheel` is called with the destination directory not
  already created

Now that the issue with invalid coverage data has been fixed.

Fix error with PEP 517 builds when wheel exists (GH #1761)

Limit workers in ProcessPoolExecutor

As a mitigation for #1730, this commit limits the number of workers in
the ProcessPoolExecutor to 1 (default is the number of CPUs). On PyPy,
having a higher number of available workers dramatically increases the
number of concurrent processes, leading to some resource exhaustion
issues.

This does not address the root issue, but should improve the situation
until the root issue is addressed.

Use xenial and latest pypy3 on Travis CI.