✨ Masking sensitive information in lang/emb requests

[roma-glushko]

Requirements

• Mask PHI, PII, and PCI information in lang/embedding requests
• Unmask all masked values before returning LLM chat responses back
• Unmask masked tokens in streaming chat requests
• Allow to enable the functionality per router
• consider using the middleware-like approach to the architecture and that there may be more pre/post processing logic to include in the future (like ✨ [Cost] Incoming Token Compression #259)

Use Case

• I want to be able to mask sensitive information (PHI, PII, and PCI) before sending it to LLM providers, so I'm not leaking it. If model returns masked placeholders, I want it to be unmasked automatically and seamlessly

[roma-glushko]

cc @mkrueger12