Fix missing .cpu() call causing WeSpeaker embedding pipeline to crash

[juanmc2005]

Problem

When waveform and masks are both on GPU, the WeSpeaker pipeline crashes because it can't convert features to numpy without moving the tensor to CPU first.

Also mentioned here: juanmc2005/diart#188

Minimal Reproducible Example

import torch
from pyannote.audio.pipelines.speaker_verification import PretrainedSpeakerEmbedding

model = PretrainedSpeakerEmbedding("hbredin/wespeaker-voxceleb-resnet34-LM")
model.to(torch.device("cuda"))

waveform = torch.randn(4, 1, 80000).cuda()
weights = torch.rand(4, 300).cuda()

emb = model(waveform, weights)

print(emb.shape)

[juanmc2005]

This is also solved by #1529
@hbredin closing this one to avoid duplicates