fix: 0.9.0 Deterministic C++ VRAM Teardown on Worker Isolate Shutdown (Issue #5)

[sebasbad]

Summary

Fixes Issue #5 on top of feat/0.9.0-mobile-first-api (PR #6):

• Previously, runEngineWorker in lib/src/isolate/worker.dart was explicitly skipping calling context.dispose(), model.dispose(), and multimodal.dispose() during ShutdownCommand.
• This left the native C++ llama_model* and llama_context* pointers and Metal GPU buffers in C++ heap memory when engine.dispose() was called, causing subsequent model load operations in long-running app processes (like mobile apps) to fail with failed to load model.
• _shutdown now explicitly disposes context, model, and multimodal before completing ShutdownCommand.

Closes #5.

[sebasbad]

Superseded by upstream PR netdur#109