Replace detector ensemble with real pYIN + cepstrum fusion

[denizyesilirmak]

• Remove YinPitchDetector entirely; pYIN is now the sole time-domain detector
• Rewrite PyinPitchDetector as real frame-wise pYIN (Mauch & Dixon 2014): Beta(2,18) threshold prior, per-candidate probability mass over all CMND minima, deepest-minimum fallback, and a light pitch-continuity bonus in place of the paper's HMM stage
• Replace EnsembleSelector voting with DetectorFusion: pYIN is the precise primary, cepstrum a coarse corroborator; frequencies are never averaged and mismatched confidence scales are never compared (agree/clash/solo rules)
• Fix cepstrum returning confident junk on pure tones: pick only strict local maxima (no range-boundary envelope tail) and use scale-invariant peak-vs-rival prominence instead of the saturating snr/5 confidence
• Drop unused IPitchDetector::setThreshold; rename cryptic variables (diff_/cmnd_ → squaredDifference_/normalizedDifference_, etc.)
• Rebuild tests around the new architecture (fusion decision tests with stub detectors, harmonic-tone cepstrum test); note tests were silently passing in Release builds with asserts compiled out