Python library for building encoderfiles

[besaleli]

Right now encoderfile can only be used via the CLI, which makes integration with Python-based ML workflows (training pipelines, notebooks, CI) awkward. We should provide a small Python library that exposes the encoderfile build functionality programmatically while preserving the current guarantees (toolchain-free, offline, deterministic). The initial implementation should be a thin wrapper around the existing encoderfile binary (shelling out, validating inputs, surfacing errors cleanly), not a reimplementation of the build logic. The API should mirror encoderfile build closely and support both structured arguments and existing YAML configs. Non-goals for v1: replacing the CLI, runtime inference in Python, or reimplementing ONNX export. Open questions include whether the Python package should manage downloading the base binary and how strict version matching should be. Success = being able to build a valid encoderfile from Python without requiring a Rust toolchain.