Langroid

Task: Restore old logic of process_invalid_step_result(): insert a DO-NOT-KNOW response if not valid response found in step.

* Improve Task infinite loop check to handle repeats like A 1 B 1 C 1 A 1 B 1 C 1 etc
* Task - do not create dummy NO_ANSWER response when step has no response (i.e. pending message does not change)
* Task - fix llm model displayed, to use agent.llm.config not agent.config.llm
* Switch examples to default to gpt-4o
* Task - when sub-task receives a msg, make deep copy, not modify orig

For ollama models, allow `api_base` to be set via OpenAIGPTConfig()

* Chainlit callback -- add option to not show sub-task response
* Task adjustments for non-interactive mode:
- do not create dummy user response when non-interactive
- if non-interactive and user not explicitly addressed in pending msg, then `can_respond(e)` returns False when `e = user`

Adds an "all" extra to get all the common dependencies.

* Improve how`ToolMessage.examples()` is used -- until now, only a random example from these examples was picked to insert a 1-shot example for the LLM. With this change, _all_ examples will be used.
* Improve pytest workflow -- first run all tests with gpt-4o, then re-run only failed ones with gpt-4