Promote Qwen3 audio encoder, S3Gen, PReLU, ESPnet subsampling, and probability segmentation to the framework

[0xShug0]

Summary

• Promote configurable Qwen3 audio encoder and S3Gen runtimes with explicit weight bindings and session-owned graph caches.
• Share PReLU, ESPnet three-stage Conv2D subsampling, and probability segmentation across their model callers.
• Reuse the existing Transformer encoder block in AST/CED and extend explicit SDPA for unequal value widths and tensor score scaling.
• Remove redundant model-local implementations and the stale Chatterbox Turbo build dependency. No ggml/kernel, server, CLI, UI, or spec changes.

Validation

• Debug CLI/server builds passed, including ten custom/core composite checks. Turbo also builds independently without Chatterbox.
• Interface-cleanup A/B: exact outputs across 50 CUDA longform requests, plus CPU/Vulkan repeated-request coverage and CUDA cloning/voice-conversion lifecycle checks.
• No persistent end-to-end slowdown established in those comparisons. Metal/HIP were not tested.
• Final CUDA interface-cleanup A/B table: report. Its baseline is identified explicitly; it is not a whole-branch-versus-main benchmark.