Fix: Add AuthenticationError guard + correct indentation in memory extensions (NVIDIA NIM users)

[MrTrenchTrucker]

Problem

Fixes #1360

For Agent Zero users running NVIDIA NIM models via LiteLLM (provider: openai with NVIDIA
api_base), an empty utility_model.api_base causes all utility model calls to route to
api.openai.com, which rejects the NVIDIA key with AuthenticationError.

Agent Zero may attempt to self-repair this by patching _50_recall_memories.py and
_51_memorize_solutions.py — but the patch can be written with incorrect indentation (1-space
instead of 8-space), causing IndentationError on extension load and blocking all agent
responses.

Changes

• _50_recall_memories.py: Fix indentation in memory_recall_query_prep block; add try/except AuthenticationError guard that detects nvapi- key on OpenAI endpoint and skips gracefully instead
  of crashing the background thread
• _51_memorize_solutions.py: Fix indentation in call_utility_model try/except block; same
  AuthenticationError guard

Testing

Verified on Agent Zero v1.x with NVIDIA NIM (qwen/qwen3.5-122b) — memory recall and memorization
now skip cleanly with a log message instead of crashing when utility_model.api_base is
misconfigured.

Full incident report and white paper:
https://github.com/MrTrenchTrucker/agent-zero-nvidia-routing-fix