Add Irodori codec backend selection

[0xShug0]

Summary

• add irodori_tts.codec_backend=same|cpu as an Irodori-only session option
• keep condition encoder and RF sampler on the requested backend while allowing the DACVAE codec decode to run on CPU
• keep old standalone GGUF packages usable when their embedded spec predates this option

Validation

• cmake --build build/debug -j 16 --target audiocpp_cli
• Vulkan default path logs irodori_tts.codec.backend same
• Vulkan + --session-option irodori_tts.codec_backend=cpu logs irodori_tts.codec.backend cpu and produces valid audio
• invalid value rejects with Invalid irodori_tts.codec_backend: bad
• Qwen3-ASR on the CPU-codec output produced largely matching Japanese text

[0xShug0]

@francescobozzo @fedeizzo It’s weird that this model-scoped change failed the Nix/Vulkan CI. Any thoughts on what might be causing this?

[francescobozzo]

Interesting, the build "worked on my machine"... I tested using my nix system installation and also reproducing github workflows with containers

git clone git@github.com:0xShug0/audio.cpp.git audio.cpp-remote
cd audio.cpp-remote
gh pr checkout 189

nix build .#vulkan

nix run nixpkgs#act -- -W .github/workflows/nix-build.yml

[0xShug0]

@francescobozzo Retry works. “Have you tried turning it off and on again?” I should’ve followed this golden rule 😄 .