SVM Subgradient Method

[hanssusilo]

Wrote SVM Subgradient script with demo function
Follows the kernel formulation
Demo runs both RBF and Linear kernels to show off SVM Sub functionality and compares to dual RLS formulation. The other methods can be easily configured in the demo script.

Requires addition of two parameters for defopt:

1. 'Niter' - the total number of iterations of subgradient descent
2. 'gammafunc' - the function that maps the iteration number to the step size

Notes:
Reuses the Dual RLS prediction code - might want to create a new one, if that suits convention better
Can be adapted to use Matlab sparse matrices to take even better advantage of SVM sparsity.