0.30.0: Install models via deeplinks

• Add Qwen 3.6 family: 27B and 35B-A3B
• Install models from Hugging Face via llamabarn:// deeplinks
• Pause and resume in-progress downloads; partials survive app quit
• Enable prompt-based speculative decoding by default
• Find sideloaded models in HF cache subdirectories; fix split-shard quant labels
• Fix MoE compatibility for sideloaded models using measured memory
• Fix sideloaded estimation hanging forever when llama-fit-params failed
• Improve sideload memory estimate accuracy
• Move models.ini to Application Support; ~/.llamabarn no longer required
• Update llama.cpp to b8902