Detect aLoRA adapters and prevent incorrect eager activation

[leehack]

Problem

llamadart currently applies every native LoRA adapter globally from the start of inference. aLoRA adapters require activation only after their invocation-token sequence. Treating them as ordinary LoRA adapters can silently produce incorrect output.

llama.cpp added aLoRA support in ggml-org/llama.cpp#15327, and the pinned b10333 bindings expose the invocation-token inspection functions.

Immediate scope

• Inspect loaded adapters for aLoRA invocation tokens.
• Reject aLoRA through a typed, actionable unsupported exception until invocation-aware execution is implemented.
• Document the current limitation so native LoRA support is not interpreted as aLoRA support.
• Add ordinary-LoRA and aLoRA/version-skew regression tests.

Follow-up/full support

• Detect the invocation-token boundary in the prompt.
• Decode the prefix with the adapter disabled and enable it at the boundary.
• Define prompt-cache behavior and prevent unsafe prefix reuse.
• Reject unsupported multiple-aLoRA combinations.
• Validate output against upstream llama-server with a real aLoRA adapter.

The safety guard should land independently of full aLoRA support.