Calling `pad_packed_sequence` with tensors on cuda and `enforce_sorted=false` causes issue

[shaltielshmid]

This bug is actually rooted in LibTorch, as reported here: pytorch/pytorch#115027

The issue is that when enforce_sorted=false the unsorted_indices are used to index on the lengths tensor, but the indices are stored on the same device as the input tensor, and lengths is stored on the CPU.

Example code to replicated:

var input = torch.randn(new[] { 5L, 5L }).@float().cuda();
var lengths = torch.full(5, 5);

var packed = torch.nn.utils.rnn.pack_padded_sequence(input, lengths, enforce_sorted: false);
var error = torch.nn.utils.rnn.pad_packed_sequence(packed);