parsing gguf head with gguf-py to avoid full loading

[excosy]

• replace optional dependency llama-cpp-python with required gguf, which can dump metadata without full loading
• remove full loading methods: analyze_gguf_with_llamacpp_tools, parse_gguf_header_simple
• add some metadata: context_length, expert_count, expert_used_count
• sync maximum of ctx_slider from gguf metadata