Apply Lee-W's review suggestion from PR apache#39007: replace the manual loop
with a list comprehension that uses ``params.dump()`` (the original
``params.items()`` iteration yielded ``Param`` objects rather than the
resolved values, which would not serialise into the Databricks API).

Extend the same pattern to:
* DatabricksRunNowOperator -> populate top-level ``job_parameters`` (the
  dict-shaped slot already supported by the run-now endpoint).
* DatabricksSubmitRunOperator -> populate dict-shaped per-task parameter
  fields (notebook_task.base_parameters, python_wheel_task.named_parameters,
  sql_task.parameters, run_job_task.job_parameters). Tasks whose only
  parameter field is ``List[str]`` (spark_jar_task, spark_python_task,
  spark_submit_task) are intentionally skipped because there is no
  canonical mapping from a key/value dict to positional CLI arguments.

Drop the ``"parameters": []`` expectation that was added to the existing
test_exec_create / test_exec_reset cases by PR apache#39007 — it never matched
the source logic (``self.params`` is falsy when no params are set, so no
``parameters`` key is added).

Add tests covering: auto-injection for each operator, no override when
the field is already populated, and the per-task injection rules for
SubmitRun.