Keep dtype and device in TextEncoder.prepend_bos_index / append_eos_index for tensor input - #3108

What does this PR do?

When TextEncoder.prepend_bos_index / append_eos_index get a tensor, the BOS/EOS index is built with torch.Tensor([...]), which is float32 on the CPU. torch.cat then promotes the whole result to float, so encoded token ids come back as floats:

import torch
from speechbrain.dataio.encoder import TextEncoder

encoder = TextEncoder()
encoder.update_from_iterable([["a", "b"]])
encoder.add_bos_eos()
tokens = torch.LongTensor(encoder.encode_sequence(["a", "b"]))
print(encoder.prepend_bos_index(tokens))  # tensor([2., 0., 1.])
print(encoder.append_eos_index(tokens))   # tensor([0., 1., 3.])

Feeding that into an nn.Embedding fails with "Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int", which is presumably why the integration tests and g2p/dataio.py call .long() on the result. The index tensor is also always created on the CPU, so a CUDA input should hit a device mismatch in torch.cat (I only have a CPU box, so I checked that with a meta tensor, not a GPU). I switched both to x.new_tensor(...) so the result keeps the input's dtype and device, and added a test that uses the device fixture. I left the existing .long() calls alone; they're harmless now.

No breaking changes unless someone relied on getting floats back. The encoder tests, its doctests and the ASR_seq2seq integration test pass locally (CPU only).

Didn't find this reported; the only other open PR touching encoder.py (#3102) just carries the limited_labelset_from_iterable change from #3095.

PR review