Publish C API libraries and add Windows Vulkan portable releases

[0xShug0]

Summary

• Add separate C API library archives for CPU, Vulkan, CUDA, and Metal, with libraries, the public header, and license.
• Add Linux CUDA 12.8/13.3 library-only builds and separate CUDA runtime archives. Windows libraries reuse the existing separate CUDA runtime bundles.
• Add Windows Vulkan portable binary and library builds with a baseline CPU target, without requiring AVX/AVX2.
• Preserve existing binary build settings, artifact contents, and Colab packaging. Library steps follow existing binary uploads; release publication also requires the new jobs to succeed.
• Locate dynamic backend plugins beside the C API library, independently of the client executable and working directory. CLI/server and static-backend initialization remain unchanged.
• No model, framework, or ggml implementation changes.

Validation

• The final workflow completed successfully in the test repository without publishing a release: artifact build.
• Downloaded Linux CPU, Vulkan, CUDA 12.8, and CUDA 13.3 libraries passed the existing C API Silero VAD and Kokoro tests; CUDA tests used the matching separate runtime archives.
• Downloaded portable libraries passed when launched from their library directory, but failed plugin discovery from an unrelated directory. Those artifacts used older source without this PR's library-relative discovery fix; updated portable archives still need rebuilding.
• The discovery fix previously passed Debug C API contract, relocated-client, export-surface, VAD, and Kokoro checks.
• Structured workflow comparison confirmed existing binary settings and steps are preserved. Windows/macOS runtime behavior remains untested.

Closes #779

[0xShug0]

Model	Peak memory reduction	Speedup
Higgs Audio TTS	48% VRAM	1.01-1.09x CUDA
ACE-Step family	6–7% total VRAM	1.06–1.08× CUDA, 1.16–1.20× Vulkan
MOSS-TTS v1.5 cloning	21% VRAM	1.05× CUDA
MOSS-TTSD Q8 cloning	11% VRAM	1.04× CUDA
Echo-TTS, memory saver	20% VRAM	---
Qwen3-TTS	16–20% VRAM	---
IndexTTS2 / 2.5, ordinary synthesis	12% VRAM	---
HTDemucs	---	2.21× CUDA, 1.95× Vulkan
HTDemucs six-stem	---	1.99× CUDA
PocketTTS	9% RAM	2.23× CPU