Fixing corner case where pad was not taken into account when decrypting the end of...

Fixing corner case where pad was not taken into account when decrypting the end of a file in a stream

__dec_read function reads directly from the file and returns the data
decrypted. This means that if the file is not encrypted, this function
is trivial.

If the data in the file is encrypted, then the process is different:
first we have to read the raw encrypted data, then decrypt it and
return. But the decryption process is not straightforward because the
self.fileobj stream contains multiple encrypted files one after the
other. We need to detect each separate file, which is detected because
they are separated by the "Salted__" keyword.

It gets more complicated, because we decrypt chunk by chunk, and to
correctly decrypt one chunk we need to set a "last" variable that
specifies if it's the last chunk of a file, because the end of a file is
handled differently, as it gets padded.

Knowing if the current chunk is the last part of a file is usually done
just by detecting if it's followed by a "Salted__" keyword or if we
cannot read more bytes from the stream. BUT there's a pretty particular
case, in which the current chunk ends exactly with one file, so that
the next chunk starts with "Salted__".

To fix that rare case, we just read N bytes from the stream, and check
if the last bytes correspond with the string "Salted__". Then we save
those last characters for next call to __dec_read. If the last bytes
were "Salted__", then we set "last" to True.

Well, actually we not only substract the length of "Salted__", but 16/32
chars because the file is decrypted in multiples of the key size.