v1.1.1

• Update prompts for GPT-5: contrary to previous models, GPT-5 attempts to
  return multiple actions in one response (not a repeated output, but intentional
  behaviour consisted with reasoning output), a line was added to prompt to
  explicitly forbid this, should not affect other LLMs.
• Verified support for GPT-5, Claude 4.5, and Gemini 2.5.
• New benchmark template execution code, now works with partial sessions with
  missing episodes.
• Template naming consistency (remove _r0 suffix).
• Minor tweaks and fixes