refactor serialize_json to use custom JSONEncoder for ~3.7x speedup

[danfunk]

Instead of two passes (Python dict-building + json.dumps), serialize_json now uses a SpiffEncoder that lets the C-level JSON encoder handle plain data traversal natively, only calling back to Python for registered types.

Periodic serialization test (300 items, 180 serializations):

Metric	Before	After	Improvement
Total time	162.5s	45.0s	3.6x faster
Total serialization	160.7s	43.3s	3.7x faster
Avg per serialization	0.893s	0.241s	3.7x faster
Final checkpoint (308 tasks)	2.07s	0.56s	3.7x faster

[essweine]

I guess this is not terrible, though all these changes are not really addressing the problem, which is that if you need to work with large amounts of data attached to each task, the library isn't the appropriate place to deal with it.