Fix multi-turn conversation ingestion for AzureOpenAI

[lindseystead]

Fixes #83

This PR fixes an issue where only the first conversation turn was being ingested
when using AzureOpenAI with Memori.

Root cause

On subsequent invocations, conversation_id was not reliably resolved early
enough in the request lifecycle. This prevented prior messages from being
injected and caused new messages to not consistently be written to the same
conversation.

Changes

• Ensured session_id and conversation_id are resolved early in
  inject_conversation_messages so prior messages can be injected correctly
• Used idempotent conversation.create() to safely retrieve or create the
  active conversation within the timeout window
• Ensured only new messages are written while injected history is excluded
  to prevent duplication
• Added logging to surface when memory is written and which conversation/session
  is used
• Added test coverage verifying all conversation turns are ingested

Result

• Every user message is ingested on every turn
• Every assistant response is ingested on every turn
• Previous messages are injected for context without duplication
• Public APIs and AzureOpenAI usage remain unchanged

[devwdave]

Hi @lindseystead
Thanks for catching this and making a contribution. Do you mind creating an issue(bug) for this PR? We will review the issue and PR and provide feedback if necessary.

Thanks again!

[lindseystead]

Created issue #233 as requested. The issue documents the bug and links to this PR.