Database persistence for spambayes classifier data could be faster

[Jean-Paul Calderone]

The database storage implementation for spambayes classifier information removes the O(N) cost of reading and writing a huge pickle, but there's still a big cost for individually loading spam/ham information about each document token individually.

It would be better to load and save all of this data in a single statement to eliminate the repeated go-into-the-database overhead.

[Jean-Paul Calderone (exarkun)]

I added a benchmark in this branch. Here are the results on my laptop:

exarkun@top:~/Projects/Divmod/branches/spambayes-fewer-potatoes$ python Quotient/benchmarks/spambayes
Learning: 0.71 ms/word
Guessing: 0.09 ms/word
exarkun@top:~/Projects/Divmod/branches/spambayes-fewer-potatoes$ chbranch Divmod spambayes-fewer-potatoes
exarkun@top:~/Projects/Divmod/branches/spambayes-fewer-potatoes$ python Quotient/benchmarks/spambayes
Learning: 0.72 ms/word
Guessing: 0.02 ms/word

I guess I shouldn't have tried to speed up writes (learning), SQLite3 must do a pretty good job of those already. Read (guessing) performance seems like a good win, though.

Also did a little bit of test enhancement so the various extra codepaths introduced by caching all get covered.