fix(history): give each MLflow run one owning task

[HackedRico]

Description

Fixes #29.

mlflow.tracking.fluent._active_run_stack is a thread-local, and every
/plugin/mcp/execute request is an asyncio task on one aiohttp event-loop
thread. Concurrent runs therefore shared one active-run pointer, so a bare
end_run() terminated someone else's run, tags landed on the wrong run, and a
write with an empty stack silently minted a phantom run.

• New app/mlflow_run.py: RunTracker binds an MlflowClient to one run id.
  Writes are truncated and best effort so observability never fails the run it
  describes; terminate() writes once.
• _run_execution and both workflows (author, plan_execute) drop the fluent
  API. A run handed in by mcp_svc is terminated by mcp_svc; a run a workflow
  mints itself is terminated by that workflow.
• hook.py runs a boot-time sweep reconciling any RUNNING run absent from the
  live cache to KILLED, which also clears existing residue.
• History Model column now has a model param to read, and Result reads the
  process_result tag instead of a param nothing ever wrote.
• History view renders KILLED and shows - for its duration, since the sweep
  timestamp is when we noticed, not when the run stopped.

Two side effects worth flagging: DSPy autolog traces are no longer attached to a
run (they still land in the experiment), and the sweep assumes one tracking
server per deployment, which is what hook.py already provisions from yaml.

Type of change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

17 new tests across tests/test_mlflow_run_lifecycle.py,
tests/test_run_execution_lifecycle.py, and tests/test_author_guards.py,
covering overlapping runs, failure attribution, phantom-run absence, run
ownership, and the orphan sweep. Each was confirmed to fail against pre-fix code
and pass after. Verified against mlflow 3.15.1 and dspy 3.3.0.

Full plugin suite compared before and after in the same environment: no new
failures, 17 new passes. Remaining suite failures are pre-existing and
environmental (missing CTI extras, live-Caldera-dependent modules).

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works