Refactor and simplify MultivariateNormal distribution

[fritzo]

This refactors and simplifies Pyro's MultivariateNormal distribution into a PyTorch-style TorchMultivariateNormal distribution plus a wrapper. This makes it easy to swap in a PyTorch MultivariateNormal as soon as it is available (as soon as probtorch/pytorch#52 is merged).

This PR is needed to support @fehiepsi 's Gaussian Process tutorial #650 as Pyro migrates to PyTorch distributions.

[jpchen]

LGTM

[jpchen]

failing tests though:

E           RuntimeError: Lapack Error in potrf : the leading minor of order 2 is not positive definite at /remote/pytorch/aten/src/TH/generic/THTensorLapack.c:617

[neerajprad]

Thanks for moving this into torch.distributions!

[fehiepsi]

@fritzo My GPRegression tutorial runs fine with this pull request. However, for a general case (when I use other likelihoods such as Bernoulli for GPClassification), torch.potrs will throw errors RuntimeError: Lapack Error in potrf : the leading minor of order ... is not positive definite at ....

Here is a simple script to replicate the problem

import torch
from torch.autograd import Variable

import pyro
import pyro.distributions as dist

K = Variable(torch.rand((1000, 1000))).double()
K = K @ K.t()

y = Variable(torch.zeros(1000)).double()

z = pyro.sample("z", dist.MultivariateNormal(y, K))

Of course, the problem is more related to "how to deal with such RuntimeError", rather than this implementation of MultivariateNormal.

[fehiepsi]

I think that it is better to raise an issue to discuss, rather than comment here.

#696