Clone all params of distribution to fix a bug

[ktaaaki]

I found a bug regarding the saving/loading Pixyz object.

1. Probrem

First, I make Normal distributions with the same tensors. (p.loc=q.loc=q.scale)

>>> from pixyz.distributions import Normal
>>> ones=torch.ones(2)
>>> p=Normal(loc=torch.zeros(2), scale=ones)
>>> q=Normal(loc=ones, scale=ones)

Next, I make new distribution with different params & load them to old one.

>>> save_path = "./tmp.pt"
>>> torch.save(Normal(loc=torch.zeros(2), scale=2 * torch.ones(2)).state_dict(), save_path)
>>> q.load_state_dict(torch.load(save_path))
<All keys matched successfully>

Then, loaded params are not correct. Inplace-updates overwrites linked params.

>>> q.scale
tensor([[2., 2.]])
>>> q.loc
tensor([[2., 2.]])
>>> p.scale
tensor([[2., 2.]])

2. Change

This is because DistributionBase class treats buffered parameters as variables of Reference types.

Therefore, I added the clone() method when registering features. Although it wastes memory, it correctly works.

And this change is related to pull request #110. Just cloning solve its issue, so I removed some lines.