feat: experiment overlays (-x flag) and benchmark correctness fixes

[jordanauge]

Adds -x / --experiment-overlay to benchmark run; fixes 3 benchmark bugs
found during development.

New ExperimentOverlay manifest kind: a YAML file that deep-merges into
experiment.yaml before validation, keeping operational concerns out of
the experiment spec.

mas-lab benchmark run experiment.yaml
-x overlays/local-dev.yaml \ # cache + n_runs override
-x pipelines/standard-eval.yaml # add reusable eval pipeline

Semantics:

• Precedence: experiment.yaml < overlay(-x) < --infra < --set
• RFC 7386 JSON Merge Patch: dicts merge recursively, everything else
  (scalars, lists) replaces. To extend a pipeline list, provide the
  full desired list in the overlay.
• Paths resolved relative to experiment.yaml directory

Files: execution/experiment_overlay.py, engine.py, run_batch/{init,
load}.py, worker.py, cli/commands/benchmark/run.py

• cache: runs with status=error in result.json are no longer treated as
  cache hits; absent or unreadable result.json is also a miss (a run
  that crashed before writing its result must not be reused)
• plotnine: force matplotlib.use('Agg') before import so the step works
  from benchmark worker threads, not only from the main thread
• bootstrap.py: remove stale import of mas.runtime.boundary.context.skills
  deleted in cb8709d; the dead import crashed every agent run

Signed-off-by: Jordan Augé augjorda@cisco.com