feat: update llama.cpp to d2a818231

[abetlen]

Updates the vendored llama.cpp revision and synchronizes the affected Python bindings.

• Advances llama.cpp from e3546c794 to d2a818231.
• Adds the new model load-mode API while preserving the existing high-level use_mmap and use_mlock options.
• Synchronizes MTMD text inputs with the new explicit byte-length field.
• Validates the CPU and MTMD build, ctypes ABI layouts, lint checks, and the full Python test suite.