llama.cpp merges Talkie-1930 model support

This is less about a flashy new model and more about ecosystem plumbing: unusual checkpoints only matter if the dominant local runtime can load them cleanly.

• –Talkie’s architecture diverges enough from standard Llama-style models that it needed its own implementation path, not just a converter tweak
• –The PR notes numerical edge cases in quantization, which is a reminder that exotic architectures can break in the places local inference users care about most
• –Support landed with GGUFs and parity testing against the official PyTorch code and HF port, so it looks aimed at practical usability rather than novelty
• –For developers, this expands the surface area of llama.cpp from mainstream chat models into historical-model research and weird-model experimentation
• –The broader signal: llama.cpp is still the default compatibility layer for whatever the community decides to run locally