refactor(pipeline): unify execution loop and document semantics (#5)

- Replace duplicated run/run_streaming/observe loops with run_pipeline_core
- Add OutputHook to inject streaming behavior
- Make root span handling uniform with fan-out spans (Span + Instant)
- Clarify carry-forward precedence, fan-out merge format, and ordering
- Tighten docs for contributors and stage result ordering semantics
- Require Send for run_streaming callback to support multithreaded runtimes

Author: cspinetta