Sync backends: Fix infinite loop and prompt echo

Initial commit: Production-ready OCR with PyTorch and GGUF backends

feat: Manually construct the prompt string instead of using `apply_chat_template` and add a `repetition_penalty` to generation arguments.

feat: Add repetition penalty to generation parameters and update default PDF scale slider value.

feat: Refactor `model.generate` parameters into `gen_kwargs` and add post-processing to remove the instruction prompt from generated text.

feat: update ggml kernels, webui components, model templates, and build configurations