Add DC offset removal and configurable fade for phrase boundaries

[leostudiooo]

Adds DC offset removal and configurable fade-in/fade-out to phrase boundaries to eliminate clicking noise.

Fixes #1762

Investigation Update

After further investigation on #1762, the root cause of clicking noise is DC offset at phrase boundaries. Note that this issue only affects some voicebanks - not all models produce DC offset.

Root Cause Analysis

Each phrase may have a small DC offset (~-56dB). At phrase boundaries, sudden changes in DC offset create step functions that appear as vertical lines across all frequencies in the spectrogram.

Solution

Implemented edge-based DC offset removal and optional fading:

1. Calculate DC offset using first/last 100 samples of each phrase (boundary regions)
2. Subtract this offset from the entire phrase
3. Apply optional short fade (0-50ms) for additional smoothing

Architecture

General-purpose implementation (per @yqzhishen's suggestion):

• Features available for all synthesis engines (Classic, Worldline, Enunu, Vogen, DiffSinger, Voicevox)
• Implemented in Renderers.ApplyPostProcessing() - centralized post-processing utility
• Applied in RenderEngine.RenderRequests() - single choke point for all rendered phrases
• Settings moved from DiffSinger preferences to general Rendering preferences
• Zero changes needed to individual renderer implementations

Results

Before (raw output):

After DC offset removal:

After DC offset removal + 10ms fade:

Features

• DC offset removal: Edge-based estimation and removal (default: off)
• Phrase fading: Configurable fade duration 0-50ms (default: off, 10ms when enabled)
• Fade curves: Linear, exponential, sine, equal-power, Hann (default: Hann)
• Both features work on cached renders (applied as post-processing)
• Available in general Rendering preferences (affects all engines)

Code Quality Improvements

• Fixed double-counting bug in DC offset calculation for short phrases
• Added input validation for fade parameters
• Normalized exponential curve to reach exactly 1.0
• Extracted magic numbers as named constants

Built and tested on macOS 26.0.1 arm64.

---

Special thanks to @yqzhishen for suggesting to make this a general-purpose feature instead of DiffSinger-only! 🙏

[stakira]

1. This sounds like a vocoder issue. The best place to solve it is the vocoder. I don't think the community vocoders produce DC offset. Is it a fine-tuned vocoder?
2. The solution in this PR is kind of weird. Should at lease use a high pass filter to process the waveform.
3. Since nobody else would need this, should probably use an option in ds vocoder config to turn it on.