fix to support int labels

[mheilman]

Integer labels get converted to numpy.int64 objects when the input is read in, and these can't be JSON-serialized. This causes run_experiment to fail when trying to write JSON results files---specifically, the descriptive statistics about the training data.

I adapted this solution for a related problem:
http://stackoverflow.com/questions/11561932/why-does-json-dumpslistnp-arange5-fail-while-json-dumpsnp-arange5-tolis

Check out this craziness:

ipdb> json.dumps(np.float64(4.0))
'4.0'
ipdb> json.dumps(np.int64(4))
*** TypeError: 4 is not JSON serializable
ipdb> 

[coveralls]

Coverage increased (+0.11%) when pulling 963531f on feature/numpy_int64_label_fix into ef8ff57 on master.

[coveralls]

Coverage increased (+0.11%) when pulling 551c201 on feature/numpy_int64_label_fix into ef8ff57 on master.

[dan-blanchard]

If you're using train_file, you shouldn't use featuresets. In fact, I thought we had it set so that it would raise an exception if you tried to use both.

[mheilman]

ah, I forgot to take that out

[coveralls]

Coverage increased (+0.11%) when pulling 45ac49a on feature/numpy_int64_label_fix into ef8ff57 on master.

[coveralls]

Coverage increased (+0.11%) when pulling 45ac49a on feature/numpy_int64_label_fix into ef8ff57 on master.

[coveralls]

Coverage increased (+0.11%) when pulling 45ac49a on feature/numpy_int64_label_fix into ef8ff57 on master.

[coveralls]

Coverage increased (+0.11%) when pulling 45ac49a on feature/numpy_int64_label_fix into ef8ff57 on master.

[dan-blanchard]

Thanks for catching this!