feat(nix): add ROCm/HIP backend support

[francescobozzo]

Summary

This PR adds the ROCm build target through the nix build system introduced in #48 . It also refactors flake.nix in multiple nix modules following llama-cpp's approach.

Changes

• Split flake into .devops/nix/ (package.nix, scope.nix, python-scripts.nix) following llama-cpp's approach
• All backend flags overridable: this is especially useful to limit the number of gpu targets for ROCm. Example: .override { rocmSupport = true; rocmGpuTargets = ["gfx1151"]; }
• python-scripts standalone package for model_manager.py
• Removed vulkan/cuda torch builds in favour of vanilla build as cpu device is enforced
• Removed x86_64-darwin as no more supported from upstream nixpkgs
• Added missing numpy dependency in python-scripts
• Updated docs/build/nixos.md

Tests

On Linux x86:

nix run .#rocm-gfx1151 -- --task tts --model /home/fbozzo/projects/audio.cpp/models/Qwen3-TTS-12Hz-1.7B-Base --backend rocm --text "audio.cpp is running locally." --family qwen3_tts --voice-ref assets/resources/sample.wav --reference-text "Some call me nature. Others call me Mother Nature. I've been here for over 4.5 billion years. 22,500 times longer than you." --out out3.wav

nix shell .#python-scripts -c python3 tools/model_manager.py install supertonic_3

On MacOS:

nix build 'git+https://github.com/francescobozzo/audio.cpp.git?ref=feature/nix-rocm#metal'

[0xShug0]

@francescobozzo Merged, thanks!