on Windows if source name ends with os.altsep.

_active shouldn't be cached, it set to None on shutdown.

This is more RFC compliant (see issue) and fixes a problem with
signature verifiers rejecting the part when signed.  There is some
amount of backward compatibility concern here since it changes
the output, but the RFC issue coupled with fixing the problem
with signature verifiers seems worth the small risk of breaking
code that depends on the current incorrect output.

Addressing #issue 20529

Issue #20540: Fix a performance regression (vs. Python 3.2) when layering a multiprocessing Connection over a TCP socket.
For small payloads, Nagle's algorithm would introduce idle delays before the entire transmission of a message.

proper CAN headers.

This applies only to the new parser.  The old parser decodes encoded words
inside quoted strings already, although it gets the whitespace wrong
when it does so.

This version of the patch only handles the most common case (a single encoded
word surrounded by quotes), but I haven't seen any other variations of this in
the wild yet, so its good enough for now.

This is a bit of an ugly hack because of the way generator pieces together the
output message.  The deepcopys aren't too expensive, though, because we know it
is only called on messages that are not multiparts, and the payload (the thing
that could be large) is an immutable object.

Test and preliminary work on patch by Vajrasky Kok.

Patch mostly by Serhiy Storchaka.

input was truncated in BASE64 section.

This fixes a regression relative to python2.

I'm checking this in without a test because not much of this code
is tested and I don't have time to work up the necessary extensions
to the existing test framework.

The patch itself was tested by the person who reported the bug.

Patch by Berker Peksag.

import and not *an* implementation.

containing non-encodable characters.  Added tests for TarFile.list().
Based on patch by Vajrasky Kok.