fix: eliminate Linux Docker crash and restore MLX on native macOS

[king-dopey]

Problem

Linux Docker containers crashed immediately on startup with:
ImportError: libmlx.so: cannot open shared object file: No such file or directory

Root cause: __init__.py eagerly imported JinaMLXReranker at module level, which triggered a top-level import mlx.core as mx before backend selection could occur. On Linux, MLX shared libraries do not exist.

A secondary regression caused MLX to no longer auto-install on native macOS after dependencies were restructured as optional extras.

Changes

src/local_reranker/init.py

• Replaced eager from .jina_mlx_reranker import JinaMLXReranker with a __getattr__ lazy-loading implementation so MLX is only imported when actually accessed, making package import safe on Linux/Docker.

src/local_reranker/config.py

• Added get_mlx_platform_error(): returns an error string when platform is not macOS arm64, None otherwise.
• Added validate_backend_type(): raises ValueError with an actionable message if mlx backend is selected on an unsupported platform.

src/local_reranker/api.py

• Moved MLX import inside the mlx branch of the lifespan startup function, after validate_backend_type() is called, so it is never reached on Linux.

src/local_reranker/cli.py

• Added validate_backend_type() call before server startup so CLI fails fast with a clear error message if mlx is requested on a non-macOS platform.

src/local_reranker/reranker_mlx.py

• Added init-time platform check via get_mlx_platform_error(); raises RuntimeError with an actionable fallback suggestion if not on macOS arm64.
• Moved JinaMLXReranker import inside _load_mlx_reranker() with try/except for ImportError/OSError.

src/local_reranker/reranker_pytorch.py

• Added warnings.filterwarnings to suppress the harmless torch_dtype is deprecated, use dtype instead warning emitted by Jina model files downloaded from HuggingFace.

pyproject.toml

• Moved MLX dependencies from [project.optional-dependencies] back to main [project.dependencies] with platform markers: mlx>=0.30.0; sys_platform == 'darwin' and platform_machine == 'arm64' This restores auto-installation on native macOS Apple Silicon without requiring a separate install step or extra flag.
• Pinned transformers>=4.57.0,<5 to avoid breaking API changes in 5.x that affect Jina model loading inside Docker.

Dockerfile

• Added RERANKER_BACKEND_TYPE=pytorch environment variable.
• CMD already hardcodes --backend pytorch ensuring the container never attempts to load MLX regardless of any env var override.

tests/test_cli.py

• Added test_main_unsupported_mlx_backend: verifies mlx backend fails with SystemExit on unsupported platforms.
• Added test_package_import_keeps_mlx_lazy: confirms jina_mlx_reranker is not present in sys.modules after a bare package import.

README.md

• Clarified that MLX auto-installs only on native macOS Apple Silicon.
• Updated Backend Support table to distinguish native macOS (supported) from Docker Desktop on Apple Silicon (not supported, Linux runtime).
• Removed outdated uv pip install -e ".[dev,mlx]" install instructions.
• Updated Troubleshooting section to reflect new dependency structure.