Make runtime model registry refresh idempotent and concurrency-safe

[amumu96]

Context

The worker update_model_type path can call register_builtin_model repeatedly while model launches continue on executor threads.

Current problems

• Built-in engine class lists use unconditional extend calls and grow on every refresh.
• Some live global model tables are cleared and rebuilt during refresh.
• Embedding and rerank derived engine entries can accumulate across repeated installs.
• Readers can observe partially rebuilt state or race with dictionary-size changes.

This behavior predates the startup engine-hook work in #5465; that PR now guarantees its callbacks run only during initial bootstrap and therefore does not address the host registry refresh itself.

Expected behavior

• Repeated registry refreshes are idempotent.
• Built-in class and derived engine entries do not grow across refreshes.
• Live readers observe a consistent registry snapshot.
• Add regression coverage for repeated LLM, embedding, and rerank refreshes.