Empty frequency sets and flat chroma profiles produce harmless warnings
during key detection. Now suppressed with targeted warning filters and
NaN handling for zero-std correlations.

…scan

Override the AudioAnalysisProvider.analyze_file hook so upstream's
AudioAnalysisController._run_background_scan can drive backfill through
the generic provider-agnostic interface. Loads audio via librosa, runs
block feature extraction and collapse, populates duration and true_peak.

Three cleanups in one commit:

1. Stop computing overlap fields in librosa:
     bpm            <- overlaid by smart_fades (beat_this CNN)
     key, mode      <- overlaid by smart_fades (S-KEY)
     danceability   <- overlaid by clap_analysis (zero-shot, calibrated)

   These were quality-inferior to their overlay sources and the overlay
   system guaranteed replacement at vector-assembly time. Computing them
   in librosa was wasted work; leaving their AudioAnalysisData fields
   None makes the architecture honest. Install must have the relevant
   overlay providers enabled or vectors won't assemble — the "no valid
   signatures found" diagnostic added in the previous commit tells the
   user exactly which fields are missing when that happens.

2. Remove dead-code feature extractions:
     librosa.feature.mfcc
     librosa.feature.tonnetz
     librosa.feature.spectral_rolloff
     librosa.feature.zero_crossing_rate

   These were extracted per block and stored on BlockFeatures but never
   read by collapse_to_analysis. Legacy from an earlier vector schema.
   Removing saves roughly 100ms per 10s block of analyzed audio — for
   a typical 3-min track, ~1.8s less CPU per analysis.

3. Fix pre-existing stale field names in test_helpers.py:
     rms_energy_per_second       -> rms_energy
     spectral_centroid_per_second -> spectral_centroid

   These referenced the pre-upstream-alignment field names and had been
   silently failing 2 tests since the AudioAnalysisData model was updated.

Net: -157/+69 lines in helpers.py, test surface shrunk to match.
All 102 sonic_analysis + sonic_similarity tests pass.

extract_block_features previously called four librosa feature functions
that each computed their own STFT internally — four redundant spectrograms
per 10s block. All four (chroma_stft, spectral_contrast, spectral_centroid,
spectral_flatness) share the same default n_fft=2048 / hop_length=512, so
a single up-front STFT is the correct input for all of them via librosa's
`S=` kwarg.

Verified byte-identical output to the old per-feature path (max abs diff
= 0 on all four feature matrices). All 10 sonic_analysis tests pass
unchanged.

Measured: 1.56x speedup on a 10s block (25ms -> 16ms). For a 3-min track
(18 blocks), that's ~180ms saved per analyzed track. At a user's 12k-track
library scale, ~36 minutes of CPU time per full background scan. Millions-
of-tracks libraries benefit proportionally.

RMS and onset_strength are left unchanged: RMS is time-domain, and
onset_strength uses a mel spectrogram with different parameters.