CPU optimization: background thread + decimation

[user1303836]

Summary

Reduces CPU usage from ~10% to <1% per plugin instance by addressing the root cause: 3x FFT-4096 running 331 times/sec on the audio thread.

Architecture change

• Pitch analysis moved to a dedicated worker thread via lock-free SPSC FIFO (juce::AbstractFifo)
• Audio thread per-sample work reduced to: 1 FIFO write + 2 atomic reads
• Worker thread handles all FFT, YIN, and decimation work off the real-time path

Algorithmic optimizations

• Hop size: 3ms -> 12ms (~4x fewer analyses/sec, from 331 to ~83)
• 2x decimation: Halfband FIR filter downsamples 44.1kHz -> 22.05kHz before analysis, halving FFT size from 4096 to 2048
• Wavetable sine: 2048-entry LUT with linear interpolation replaces std::sin() per sample

Code-level optimizations

• Two-memcpy buffer linearization (eliminates per-element modulo in ring buffer read)
• FloatVectorOperations::clear for SIMD-accelerated FFT buffer zeroing
• Cached exp2 in PitchSmoother (recomputes only when smoothed value changes)
• Modulo replaced with branch in ring buffer write

Files changed

File	Change
source/dsp/HalfbandDecimator.h	New: 7-tap halfband FIR decimation filter
source/dsp/YinPitchDetector.h	FIFO, thread, atomic members; inline feedSample/getResult
source/dsp/YinPitchDetector.cpp	AnalysisThread inner class, decimation, code-level fixes
source/dsp/PitchSmoother.h	Cached exp2 result
source/dsp/Oscillator.cpp	Wavetable sine lookup
tests/TestYinPitchDetector.cpp	flushForTest() for async, updated timing helpers

PluginProcessor.cpp is unchanged -- the public API (prepare, feedSample, getResult) is preserved.

Test plan

• All 34 tests pass (441,095 assertions)
• Full plugin builds (VST3)
• pluginval passes at strictness 5
• CI passes on macOS and Windows
• Manual test: confirm pitch tracking works identically in DAW