Expand FFI bindings with grammar, LoRA, embeddings, and more

[nbeerbower]

New FFI bindings:

• Grammar sampling (llama_sampler_init_grammar) for structured output
• LoRA adapter support (load, apply, remove adapters)
• State/session save and load functions
• Additional samplers (XTC, top-n-sigma, infill, DRY)
• FIM (Fill-in-Middle) tokens for code completion
• Extended model metadata access
• Chat template formatting
• Embeddings extraction helpers
• Control vectors support
• Context settings (causal attention, warmup mode)
• Additional vocab functions (sep, mask, score, attr)
• NUMA and RPC support detection

New high-level API functions:

• sampler_create_with_grammar() - Grammar-constrained sampling
• lora_load/apply/clear() - LoRA adapter management
• state_save/load() - Session persistence
• chat_format() - Format messages with model template
• extract_embeddings() - Get text embeddings as array
• fim_prompt() - Create FIM prompts for code models
• model_info_extended() - Extended model information
• model_meta/model_meta_all() - Access GGUF metadata

New examples:

• json_output.hml - Generate structured JSON
• lora.hml - LoRA adapter usage
• embeddings.hml - Text embeddings and similarity
• fim.hml - Fill-in-middle code completion
• session.hml - Save/restore conversation state

Constants added:

• Token types and attributes
• ROPE_SCALING_LONGROPE
• GRAMMAR_JSON (built-in JSON grammar)