Refactor pipeline discovery to use Python entry points, enabling external
pipelines to live in separate repositories and be installed as dependencies.

Key changes:
- Created pyproject.toml for ai-runner-base package with entry points
- Refactored loader.py to discover pipelines via entry points (no hardcoded imports)
- Added two entry point groups: ai_runner.pipeline and ai_runner.pipeline_params
- Removed Params class attribute from Pipeline interface (params discovered via entry points)
- Updated Dockerfiles to use pyproject.toml for dependency installation
- Removed requirements.live-ai.txt (deps consolidated into pyproject.toml)
- Updated all documentation with plugin system details
- Added Docker image hierarchy documentation to AGENTS.md

Benefits:
- Pipelines can live in separate repos and be installed via pip/uv
- No code changes needed in ai-runner when adding new pipelines
- Params can be parsed without importing heavy dependencies (torch, etc.)
- Works seamlessly with multiprocessing spawn context
- Cleaner separation of concerns between runtime and pipeline implementations