fix: automatic memory injection for Google GenAI provider (#215)

[KarthikeyaKollu]

fixes #215 This commit fixes two critical issues that prevented memory injection from working with the Google GenAI SDK:

1. _extract_user_query now handles Google GenAI's 'contents' parameter:

   • Supports string contents (simple query)
   • Supports list of strings
   • Supports structured contents format with parts/role

2. inject_recalled_facts now properly injects context for Google GenAI:

   • If config.system_instruction is available, appends context there
   • Otherwise, prepends context to the first user message's content
   • If first message is not a user role, inserts a new context message
   • Properly normalizes contents format (string/list/structured)
   • Ensures kwargs['contents'] is properly updated

Also fixed a minor bug where kwargs['messages'] wasn't being assigned back in the else branch for non-Anthropic/Bedrock providers.

Includes comprehensive tests for all Google GenAI scenarios:

• String contents
• List of strings
• Structured contents with parts/role
• system_instruction config
• Edge case: first message not user role