Deprecation of stateful decode for sliding-window layers in ggml-org/llama.cpp

The release notes for ggml-org/llama.cpp version b10981 announce significant changes to the stateful decode mechanism, particularly for sliding-window layers. The stateful decode now refuses to seed its KV state from ggml's cache when the decode position exceeds what the state holds. This restriction arises because sliding-window layers only retain the last n_swa positions, making the seeded state invalid beyond the window. Additionally, slicing the state to the decode position lacked bounds checking, which previously resulted in unhandled exceptions. The updated implementation now rejects these cases with clear error messages and applies the same refusal logic during compilation for new models starting with an empty state. Operators can reproduce these scenarios using llama-bench with the `-d` option, which restores a saved sequence state instead of recomputing the depth prefill.

Before upgrading, operators should verify whether their workflows rely on stateful decode functionality for sliding-window layers or involve models with empty initial states. Any pipelines or tools that depend on the previous behavior may encounter compatibility issues or outright failures. This change aligns with a broader trend in AI frameworks to enforce stricter validation and error handling, ensuring more predictable behavior but potentially requiring adjustments in legacy workflows.