This is split off a large patch I posted at Issue 20132. My new patch here fixes the following two flaws.

1. There is special code in the bz2 decoder that returns an empty text str object at EOF, even though bz2-codec is a bytes-to-bytes codec:
>>> import codecs
>>> decoder = codecs.getincrementaldecoder("bz2")()
>>> decoder.decode(codecs.encode(b"data", "bz2"))
b'data'
>>> decoder.decode(b"", final=True)  # Should return bytes object
''

2. The base64 decoder does not handle partial sets of four codes, because it treats each input chunk as a stand-alone base64 encoding:
>>> tuple(codecs.iterdecode((b"AA", b"AA\r\n"), "base64"))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/lib/python3.5/codecs.py", line 1039, in iterdecode
    output = decoder.decode(input)
  File "/usr/lib/python3.5/encodings/base64_codec.py", line 35, in decode
    return base64.decodebytes(input)
  File "/usr/lib/python3.5/base64.py", line 554, in decodebytes
    return binascii.a2b_base64(s)
binascii.Error: Incorrect padding