I have simplified the PR. Removed the complex code for detecting pickles written to files in text mode on Windows and for adding optional marks for correct detecting. Currently it does only two things:

* Escapes \r, \0 and \x1a (end-of-file on Windows) when pickle unicode. This allows pickles dumped to a file in binary mode (or on non-Windows, or in Python 3) be correctly loaded from a file opened from a file in text mode on Windows.

Currently, dumping to a file in text mode works most time, except on Windows, when the unicode string ends with \r or contains \x1a (not sure about \0, it was added just for the case). Since the data is only corrupted in special cases, this likely is not tested, and the user code can open files in text (default) mode without noticing a bug, until once a malicious user will provide a bad Unicode string.

* Produces a deprecation warning or even a ValueError in cases when writing to a file in text mode can cause problems. This will help to notice the problem earlier.