Map ClearerVoice MossFormer2_SR_48K compatibility feasibility

[Tinnci]

Follow-up from #23 and docs/SPEECH_BACKEND_CANDIDATES.md.

Goal: decide whether ClearerVoice MossFormer2_SR_48K can become a self-contained compatible backend in this package.

Scope:

• Inspect required HF files for alibabasglab/MossFormer2_SR_48K and decide which are needed for inference.
• Download to a temporary local cache only; do not commit weights.
• Inspect checkpoint structure and tensor key layout.
• Map upstream config, preprocessing, STFT/mel parameters, generator/vocoder components, sample-rate assumptions, and chunking behavior.
• Verify whether CPU inference works through upstream ClearVoice on a short fixture.
• Produce a proposed ModelSpec, manifest file list, and validation plan if feasible.

Out of scope:

• Implementing clearvoice-mossformer2-sr-compat.
• Adding large weights to the repository.
• Hardware acceleration provider work; that belongs to v0.4.0.

Acceptance criteria:

• A concise feasibility conclusion: accept for implementation, defer, or reject.
• Required files and hashes are documented if accepted.
• Known blockers are documented if deferred/rejected.