Add asymmetric hybrid block-causal mask for efficient vision path

Clean up inference config: remove training-only flags, set bd_size=32 default, dtype=bfloat16

Upload folder using huggingface_hub

initial commit