Fix three issues in Kiro OpenAI translator that caused "Improperly formed request"
errors when processing LiteLLM-translated requests with tool_use/tool_result:
1. Skip merging tool role messages in MergeAdjacentMessages() to preserve
individual tool_call_id fields
2. Track pendingToolResults and attach to the next user message instead of
only the last message. Create synthetic user message when conversation
ends with tool results.
3. Insert synthetic user message with tool results before assistant messages
to maintain proper alternating user/assistant structure. This fixes the case
where LiteLLM translates Anthropic user messages containing only tool_result
blocks into tool role messages followed by assistant.
Adds unit tests covering all tool result handling scenarios.
