Add log_prob

[masa-su]

• remove the log_likelihood function
• add log_prob in Distribution API which performs as the instance of Loss API.
• rename the LossExpectation class to Expectation and add the expectation method in Loss API.
• rename the estimation method in Loss API to eval (according to model.eval() on PyTorch).
• remove flow implementations for now.

[masa-su]

Example code

>> print(q)
Distribution:
  q(z|x) (Normal)
Network architecture:
  Inference(
    (fc1): Linear(in_features=784, out_features=512, bias=True)
    (fc2): Linear(in_features=512, out_features=512, bias=True)
    (fc31): Linear(in_features=512, out_features=64, bias=True)
    (fc32): Linear(in_features=512, out_features=64, bias=True)
  )

>> print(q.log_prob())
log q(z|x)  # log-probability density function（Loss API）

>> samples_dict = q.sample({"x": _x})
>> loss = q.log_prob().mean().eval(samples_dict)
>> print(loss)
tensor(-67.4651, device='cuda:0', grad_fn=<MeanBackward1>)

>> loss = ((p * prior).log_prob() - q.log_prob()).expectation(q).mean()  # = Expectation(q, (p * prior).log_prob() - q.log_prob()).mean()
>> print(loss)
mean(E_q(z|x)[log p(x,z) - log q(z|x)])