Implement basic Task serialization

[philippjfr]

The goal for Lumen is to have a persistence layer so explorations can be exported, shared and reloaded. Towards that goal every piece of an exploration needs to be exportable. This PR sets the foundation by implementing a consistent to_spec/from_spec protocol for Task types. This is inline with the general serialization protocol used by other Lumen components.

The default serialization includes the task tree definition and history, but excludes rendered views and generated artifacts (e.g. out_context, live UI state). The ActorTask deserialization resolves its actor via an explicit lookup mapping supplied to from_spec, i.e. the actual execution layer is distinct from the serialized state so an exported Task can be loaded into any session.

Example

A serialized Plan might look like this:

{
  "type": "lumen.ai.coordinator.base.Plan",
  "title": "State capacity",
  "history": [
    {
      "role": "user",
      "content": "Calculate total capacity by state"
    }
  ],
  "tasks": [
    {
      "type": "lumen.ai.report.ActorTask",
      "title": "Aggregate by state",
      "instruction": "Query the windturbines_parquet source and compute total wind turbine capacity grouped by US state (return state and summed capacity).",
      "actor": "SQLAgent"
    },
    {
      "type": "lumen.ai.report.ActorTask",
      "title": "Check results",
      "instruction": "Validate that the aggregation returns one row per state and that capacity values are summed correctly (no missing/duplicate states due to join issues).",
      "actor": "ValidationAgent"
    }
  ]
}