Remove .double from perplexity_compute

[peter-tk-tran]

Summary:
MPS requires F32 but .double casts the output back to F64 which causes it to fail

Example:

import torch
from torcheval.metrics.text import Perplexity

device = torch.device("mps")
logits = torch.Tensor([[[1, 9], [2, 3]]]).to(device)
labels = torch.Tensor([[0, 1]]).int().to(device)

perplexity = Perplexity().to(torch.float32).to(device)
perplexity.update(logits, labels)
print(perplexity.compute())

  File "/Users/petertran/torcheval/torcheval/metrics/text/perplexity.py", line 121, in compute
    return _perplexity_compute(self.sum_log_probs, self.num_total)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/petertran/torcheval/torcheval/metrics/functional/text/perplexity.py", line 117, in _perplexity_compute
    return torch.exp(sum_log_probs / num_total).double()
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.