Support dynamic LiteRT-LM LoRA adapters and web parity

[leehack]

Summary

Split from #173 after initial native LiteRT-LM load-time LoRA support landed.

Current llamadart support covers one default-scale text LoRA adapter supplied through ModelParams.loras when loading a native .litertlm model. The remaining parity work is dynamic/runtime LoRA control and web support.

Remaining work

• Support runtime setLora, removeLora, and clearLoras for native LiteRT-LM contexts when the runtime exposes stable APIs for dynamic adapter changes.
• Support adapter stacking and custom scales if the LiteRT-LM runtime defines compatible semantics.
• Support LiteRT-LM web LoRA if @litert-lm/core exposes equivalent browser runtime APIs.
• Keep unsupported combinations failing early with clear errors until runtime support exists.
• Add regression tests and at least one real-model smoke once test artifacts are available.

Current blocker

This needs upstream/native runtime API support beyond the current session-config lora-path setter. Dart should not bind to mangled C++ internals.

Related

• Parent: Add LiteRT-LM LoRA adapter support #173
• Current docs describe the supported load-time subset and rejected runtime/web paths.