Llama Prompt Guard 2 models ship; search_settings adds images

Changed Python SDK updated to v0.26.0, TypeScript SDK to v0.23.0. search_settings now includes include_images field for agentic tooling systems; added code_results (png/text) to executed tool outputs.

Added Meta Llama Prompt Guard 2 models (22M and 86M parameters) — specialized classifiers to detect and prevent prompt attacks (injections, jailbreaks) as part of Meta's Purple Llama initiative.

Performance (22M): 99.8% AUC English jailbreak detection, 97.5% recall at 1% FPR, 81.2% attack prevention rate. Performance (86M): 99.5% AUC, 88.7% recall at 1% FPR, 78.4% attack prevention rate, 75% latency reduction vs larger models.