project: bump to 1.7.0

[zhouwg]

Changes

• sync with upstream llama.cpp and whisper.cpp
• sync with project ggml-hexagon(v0.99.4)
• decouple ggml-hexagon.cfg config file from runtime libpath: config file is now copied to /data/local/tmp/ (user editable via adb shell), DSP skeleton libraries remain in app data directory
• update built-in AI model list: remove Qwen3-4B, add ASR model ggml-tiny.en-q8_0 to README, mark Gemma-4-E2B as default LLM model
• refine GitHub Actions CI: add APK output diagnostic step, consolidate to single qcom build target
• refine logic of AI model download
• improve stability
• refine docs
• remove stable-diffusion due to limited resource
• remove -mcpu=cortex-x1 -mtune=cortex-x1 for better compatibility with Snapdragon 8 Gen 2 & 8 Gen 3
• remove ncnn due to limited resource
• reorganize source directory structure

TODO

• unidentified issue exists in real-time video inference and will be fixed in the next PR
• unidentified issue exists in Gemma3-4B inference and will be fixed in the next PR