Add embedding-based semantic detection (beyond regex)

[manja316]

Regex patterns catch known attack strings but miss semantic equivalents:

• "Please disregard your prior directives" (same intent, no pattern match)
• Paraphrased jailbreaks
• Novel attack vectors

Consider an optional lightweight embedding similarity check using a small local model (e.g. sentence-transformers). Should remain zero-dep by default — embedding mode as optional extra.

pip install ai-injection-guard[semantic]