Memory leaks in dataset training

[alastair]

During dataset training, when cluster_mode=False, the multiprocessing code starts n new processes, and then runs jobs in each process. There is a memory leak during this process, and so eventually the processes get killed (and turn into zombies).
Running with cluster_mode=True, a new process is started for the actual training process, and so any leaked memory is recovered at the end of each job.

Attached is a valgrind report of potential memory leaks. Gaia was built with debug mode, and so symbol names and filenames are included (search for gaia2::)
gaia-svmclassification-valgrind.txt

[alastair]

gaia/src/3rdparty/libsvm/gaiasvmutils.cpp

Lines 121 to 122 in 3c07f6f

	prob.y = new double[nsegments];
	prob.x = new svm_node_ptr[nsegments];