[Feature Request] pt: support virtual types

[njzjz]

Virtual types are not supported in pt; thus, mixed-natoms input is not supported.

For example, given the following atomic types:

frame 0: 0 0 0 0 0
frame 1: 0 0 0 -1 -1

The atomic type of the second frame is equivalent to [0, 0, 0].

The current behavior is that an error is thrown:

E             File "/home/jz748/codes/deepmd-kit/deepmd/pt/model/descriptor/dpa1.py", line 196, in forward
E                   nframes, nloc, nnei = nlist.shape
E                   nall = extended_coord.view(nframes, -1).shape[1] // 3
E                   g1_ext = self.type_embedding(extended_atype)
E                            ~~~~~~~~~~~~~~~~~~~ <--- HERE
E                   g1_inp = g1_ext[:, :nloc, :]
E                   g1, g2, h2, rot_mat, sw = self.se_atten(
E             File "/home/jz748/codes/deepmd-kit/deepmd/pt/model/network/network.py", line 568, in forward
E               
E                   """
E                   return self.embedding(atype)
E                          ~~~~~~~~~~~~~~ <--- HERE
E             File "/home/jz748/anaconda3/lib/python3.10/site-packages/torch/nn/modules/sparse.py", line 163, in forward
E               def forward(self, input: Tensor) -> Tensor:
E                   return F.embedding(
E                          ~~~~~~~~~~~ <--- HERE
E                       input, self.weight, self.padding_idx, self.max_norm,
E                       self.norm_type, self.scale_grad_by_freq, self.sparse)
E             File "/home/jz748/anaconda3/lib/python3.10/site-packages/torch/nn/functional.py", line 2237, in embedding
E                   # remove once script supports set_grad_enabled
E                   _no_grad_embedding_renorm_(weight, input, max_norm, norm_type)
E               return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
E                      ~~~~~~~~~~~~~~~ <--- HERE
E           RuntimeError: index out of range in self