Improve error handling in Event Processor by aggregating multiple LLM failures

Currently, the EventProcessor only captures the first error encountered during an invocation. Any subsequent errors are ignored, which leads to incomplete diagnostic information when multiple issues occur.

This issue proposes:

Merging multiple LLM error messages into a single, concatenated string.
Including the full error message in the terminal event's metadata.
Ensuring task actions (escalate/transfer) are preserved even when errors are reported.

This follows the TODO left in processor.go:
// TODO(yarolegovich): consider merging responses if multiple errors can be produced during an invocation