Keep the LoRA runtime warm between jobs

[j316chuck]

• Add opt-in engine flags to prewarm inference and retain the shared Ray, worker, inference, and base-model runtime after the final LoRA unload.
• Unload each job's LoRA and optimizer state while allowing the next compatible create_model to register a fresh adapter against the warm runtime.
• Keep cold teardown as the default and document the multi-tenant lifecycle.

Testing

	Cold	Warm
Before	102.512s	1801.271s (failed)
This PR	100.815s	0.152s