between 2**31 and 2**63 on systems with a 64-bit C "long".

Reported by Oleg Oshmyan in issue #12646.

This will make it easier to detect truncated input streams.

Also, make zlib's error messages more consistent.

normalization is different between unicode builds, so use a new non-BMP char and add normalization test

This fixes the use of Unicode identifiers outside the BMP in narrow builds.

The tests now have two convenience functions to wrap os.path.join, open
and read or write instead of four or six slightly different functions.
The new functions accept a tuple of path segments but not a list
anymore, as it makes no sense to use a list here; I have also removed
the default value for the contents in write_file, as I find it better to
have the contents at the call site.

For simple open then read/write calls, I have left the usual idiom (with
open + read/write), as it is short and readable enough.

I’ve also changed some convoluted cleanup code to just use rmtree, and
removed dubious LBYL os.path.exists checks.  The tests still pass on my
machine, and leave no file in $TMP.  test_shutil is not as clean as it
could be, but I’ll stop here.

Initial patch provided by Hynek Schlawack, in preparation for a new
feature with new tests in #12715.

Changes to crlf originally by Victor Stinner for 3.3, copied to lfcr by
me.  Manually tested.

The line about the test suite will still get printed for changes in
Tools for example, which aren’t covered by the test suite, but it’s not
a big deal IMO.

Patch by Francisco Martin Brugue.