Add HuggingFace WavLM backend to speaker similarity metric

[NewGamezzz]

Why

Most recent TTS and voice-conversion papers report speaker similarity (SIM) using HuggingFace speaker-verification models, most commonly microsoft/wavlm-base-sv (or its variants, e.g., wavlm-base-plus-sv). Since the speaker metric only supports ESPnet-SPK models, VERSA scores cannot be directly compared against the numbers reported in those papers. This PR adds support for HF x-vector models so that users can reproduce the evaluation setup with the same config format as before:

- name: speaker
  model_tag: microsoft/wavlm-base-sv

File Changes:

• versa/utterance_metrics/speaker.py: added resolve_speaker_backend(), which auto-detects the backend from model_tag. A thin HFSpeakerModel wrapper mirrors Speech2Embedding's call signature so both backends share the existing cosine-similarity code. No new dependencies: transformers was already in the ml extra and the import is guarded.
• versa/scorer_shared.py: made the shared cache namespace config-aware so HF speaker models cache under huggingface/ instead of espnet_model_zoo/.
• versa/config_validation.py: dependency validation now only requires the selected backend's package, so a transformers-only install can run WavLM configs without being asked to install espnet2 (and vice versa).
• egs/separate_metrics/spk_similarity_wavlm.yaml (new) and docs/supported_metrics.md: example config and docs for the new backend.
• test/test_metrics/test_speaker.py and test/test_pipeline/test_speaker_wavlm.py (new): unit tests for backend resolution, caching, and validation.