llama.cpp Release b11435 Fixes CPU Data Race and Refactors Sequence Handling

llama.cpp release b11435 resolves a data race condition on CPU backends occurring during k-pool scatter operations on shared sequences. It also refactors sequence copying (seq_cp) to enforce whole-sequence operations and removes the legacy k-pool cache_safe mode. Fixing this race condition enhances stability and execution reliability when running multi-sequence inference on CPU architectures. Additionally, simplifying sequence state copying removes redundant staleness checks and overhead during context management. Instead of re-pooling shared cells across scatter entries, each shared k-pool rep is now marked only once per micro-batch. Furthermore, partial sequence copies are now explicitly rejected in hybrid memory, allowing the removal of complex stale-all workarounds across seq_rm, state_read, and state_drop.