feat(grpo): add pluggable action format (DSL vs JSON)

[abrichr]

The GRPO trainer hardcodes DSL format (CLICK/TYPE/WAIT/DONE) in prompt construction, parsing, and formatting. External RL training use cases need JSON format ({"type": "click", "x": 0.461, "y": 0.021}).

Introduce an ActionCodec protocol with encode/decode/build_prompt methods and DSL/JSON implementations.

Affected functions:

• _build_agent_messages
• _parse_vlm_output_to_action
• _format_action_as_text