Test additional losses and move them to a meaningful place

[einbandi]

Currently kullback_leibler_divergence and cross_entropy_loss are located in paradime.loss. Should they be moved to paradime.routines instead? What about the following losses from the old ptsne.py file:

def kullback_leibler_loss(p, q, eps=1.0e-7):
    eps = torch.tensor(eps, dtype=p.dtype)
    kl_matr = torch.mul(p, torch.log(p + eps) - torch.log(q + eps))
    kl_matr.fill_diagonal_(0.)    
    return torch.sum(kl_matr)

def kullback_leibler_reverse_loss(p, q, eps=1.0e-7):
    return kullback_leibler_loss(q, p, eps)


def jensen_shannon_loss(p, q, eps=1.0e-7):
    m = 0.5*(p+q)
    return 0.5*kullback_leibler_loss(p, m, eps) + 0.5*kullback_leibler_loss(q, m, eps)

def frobenius_loss(p, q):
    return torch.pow(p - q, 2).sum()

def total_variational_loss(p, q):
    return torch.abs(p-q).sum()