unprocessed items in batch write may be double encoded

[danielhochman]

now that we fixed the parser in 1.5.0 for UNPROCESSED_ITEMS in #114, it appears we are double encoding in this retry: https://github.com/jlafon/PynamoDB/blob/da6aa1068d5a496ed25bcb2b0864b4c894ffc0ce/pynamodb/models.py#L120

probably we need to decode in the parser in addition to _convert_binary_attr

[danielhochman]

binary attributes are double encoded by default unfortunately see #127

it turns out that binary attributes written by the unprocessed items flow were n+2 times encoded, where n is the number of times the item was returned as an unprocessed item from subsequent retries.