feature llm: few-shot and new algorithm registration

[babou7063]

Summary

• Add few-shot support to LLMClassifier: new few_shot_count parameter samples labelled demos from the training split and injects them into the prompt via a {few_shot_examples} placeholder.
• Add Mistral-LLM-FewShot algorithm in algorithms.yml
• Extend PromptStrategy.parse into parse_with_meta returning metadata (label, reason, p_packed, decision, fallback flag), old parse preserved as wrapper.
• Add generate_with_logprobs to LLMBackend for first-token log-probability extraction with fallback for llama.cpp builds without logprobs.
• Add FeatureFormatter representation styles, probability.py utility, types.py shared types.
• Add trace/debug output support (trace_outputs, trace_dir params).
• Fix model cache dir path (~/.packing-box/cache/llm-models).

All modifications confined to pboxllm/