fix(distributed): initialise the DDP group on the active accelerator - #3116

Problem

ddp_init_group() assumed the accelerator is CUDA:

• the rank -> device guard was sized with torch.cuda.device_count(), and
• the backend whitelist was nccl/gloo/mpi.

On a host whose accelerator is not CUDA (Ascend NPU through torch_npu, whose collective library is
hccl), torch.cuda.device_count() is 0, so every rank aborted with

ValueError: Killing process
Not enough GPUs available!

before init_process_group was ever reached — even with 4 NPUs visible — and hccl was rejected as a
backend that "doesn't exist" even though PyTorch reports it as available. This is the same device-agnostic
migration that infer_device() went through in #3080; the distributed launch path was left behind.

What

• the device count for the guard, and the per-rank device binding, now come from
  torch.accelerator (current_accelerator().type, device_count());
• the nccl/gloo/mpi checks stay as they are, and any other backend that
  torch.distributed.is_backend_available() reports as available is accepted (accelerators other than
  CUDA ship their own collective library, e.g. hccl for Ascend NPU).

CUDA hosts keep the previous behaviour: backend nccl and torch.cuda.set_device(device).

Verification

Real hardware: Ascend 910B host (aarch64, 4 NPUs, torch 2.15.0.dev, torch_npu 2.15.0.dev,
torch.accelerator.is_available() == True, current_accelerator() == npu, device_count() == 4,
torch.cuda.is_available() == False). speechbrain/utils/distributed.py from develop and from this
branch was loaded by file path on that host, and ddp_init_group({"distributed_backend": "hccl"}) was run
through torchrun --nproc_per_node=2:

develop: [rank 0] FAILED ValueError: Killing process / Not enough GPUs available!
         [rank 1] FAILED ValueError: Killing process / Not enough GPUs available!
branch : [rank 0] ddp_init_group OK | pg backend=hccl world_size=2
         [rank 1] ddp_init_group OK | pg backend=hccl world_size=2
         [rank 0] all_reduce sum=6.0 on npu:0   (two elements per rank, values rank+1)
         [rank 1] all_reduce sum=6.0 on npu:1

The repository's own test file was run on the same host against the real module (pytest 8.3.2, with the
package skeleton in place): 8 passed with this change. With the develop module and the three new tests,
2 failed, 1 passed — the failure is the "Killing process" guard above.

Not tested

• A real CUDA/nccl host — the code path is written as before (torch.cuda.set_device) but was not executed
  on a GPU machine.
• The mpi backend.
• End-to-end speechbrain recipe training: that host has torch/torch_npu but not speechbrain's runtime
  dependencies, so the real module and the repository's own tests were driven instead.