Optimize realtime silence path and SOLA hot-loop to reduce CPU/GPU overhead

[blaisewf]

Motivation

• Reduce repeated heavy pipeline runs during silence to lower CPU/GPU usage and reduce wake-up latency spikes in realtime conversion.
• Remove per-chunk allocations and temporary tensors in the SOLA/crossfade hot path to improve throughput and consistency.

Description

• Add a throttled keepalive interval controlled via keepalive_ms in **kwargs (default 200 ms) and convert it to keepalive_interval_s, tracking last call with last_keepalive_time.
• Introduce _run_silence_keepalive(...) helper that centralizes the periodic background pipeline.voice_conversion(...) call and is invoked from VAD and low-volume early-return paths.
• Cache a device-side silence tensor self._silence_template to avoid allocating zeros on each silent chunk and return it for silent fast-paths.
• Precompute and reuse sola_norm_kernel and onset_fade_window in VoiceChanger.generate_strength and replace per-request torch.ones(...) and torch.linspace(...) with these cached tensors in the SOLA denominator and onset fade logic.

Testing

• Ran python -m py_compile rvc/realtime/core.py and it completed successfully.

---

Codex Task