kv cache slot search improvements (#3493)
* kv cache slot search improvements
* Use n_ctx in kv find slot for consistency
* Ensure kv cache head points to a valid slot in llama_decode internal
* Add some comments to prevent dumb people (like me) from getting confused.
