Split from PYTHON-1374.

Context

Currently, the JSON encoder performs a chain of isinstance checks sprinkled in with a "_type_marker" lookup. We can optimize this further by using a single dispatch table similar to how the pure python BSON encoder works:

1. call type(obj) and lookup the type in the encoding table, if the type is not found continue to step 2:
2. access obj._type_marker, if it exists and is one of our known types, add type(obj) to the encoding table and encode it, otherwise continue to step 3
3. perform the chain of isinstance checks, if it's a subclass of the known type, add type(obj) to the encoding table and encode it, otherwise raise TypeError.

Definition of done

Make the above refactor and measure the performance to ensure there are no regressions and perf has improved.