Instantiating many small RawBSONDocuments is inefficient

Starting in 3.4 we can retrieve raw BSON documents from a cursor:

options = CodecOptions(document_class=RawBSONDocument)

for doc in db.get_collection("collection", codec_options=options).find():

    # doc is a RawBSONDocument.

    doc.raw

This is mainly intended for speed in libraries like Python-BSONJS that can decode raw BSON themselves. If the documents are small, however, each instantiation of a RawBSONDocument is much more costly than decoding the BSON to a dict would have been. The main penalty is from copying the CodecOptions class for each RawBSONDocument.

To avoid copying, we will have to require that the CodecOptions passed to a RawBSONDocument has document_class of RawBSONDocument. In normal usage like the example above, that is always true. But, there are direct usages of RawBSONDocument that will be prohibited now:

# Now prohibited, document_class defaults to dict.

document = RawBSONDocument(

    bson_string,

    codec_options=CodecOptions(uuid_representation=JAVA_LEGACY))

# Do this instead.

document = RawBSONDocument(

    bson_string,

    codec_options=CodecOptions(uuid_representation=JAVA_LEGACY,

                               document_class=RawBSONDocument))