Speaker similarity fix

[ftshijt]

Summary

• Add warnings when speaker similarity, singer identity similarity, and EMO2VEC emotion similarity receive audio below their embedding model target sample rates.
• Keep existing resampling behavior intact while making potentially unreliable upsampled embedding scores visible to users.
• Add lightweight tests that assert the warning behavior without requiring model downloads.

Motivation

Issue #83 notes that speaker embedding similarity can be misleading for audio sampled below 16 kHz, especially with the default RawNet3-style path. The same concern applies to similar learned embedding comparisons that silently upsample lower-rate audio before scoring.