fix: reject LLM calls with incompatible contexts

[mikemikimike]

Summary / Problem

Closes #33. llm_text_generate() and llm_embed_generate() only checked whether a context existed, so calling one with the other operation's context could silently return an empty result or run against incompatible state.

Changes

• Track whether each llama context is for text generation, embeddings, or chat.
• Reject text/embedding generation calls when the active context kind does not match, returning SQLITE_MISUSE with an actionable error.
• Add a regression test covering both mismatch directions.

Tests

• git diff --check — passed.
• Docker-based C test compilation was attempted with Alpine Linux, but the container's package installation stalled without output in the available environment; the C test binary and full test suite were not run.

Compatibility / Known limitations

Valid calls retain their existing behavior. This changes an invalid context/operation combination from a silent result to an explicit SQLITE_MISUSE error. The full model-backed test suite remains to be run in an environment with the repository's native toolchain and test model dependencies.