feat(purejava): add deterministic execution planning

[bsbodden]

Summary

• expose backend-neutral execution diagnostics and stable optimization decisions
• build one pure-Java plan from actual GGUF layer topology, JVM identity, and structured Vectors capabilities
• route grouped projections and batched prefill through the immutable load-time plan
• reject malformed overrides, contradictory plans, and plans for different weights
• expose diagnostics through LangChain4J and ModelJars profiles through Spring Boot
• document the controlled HotSpot/Graal full-model matrix

Controlled result

Best measured per-model JVM selection raises summed decode from 21.7% to 30.9% of llama.cpp and from 47.1% to 67.1% of Ollama. All compiler-paired output hashes matched.

Tests

• ./gradlew :models-api:test :models-backend-purejava:unitTest
• ./gradlew :models-langchain4j:test :models-spring-boot-starter:test
• ./gradlew check