Draft: Add LRU cache for model.bind_tools to resolve CPU bottleneck

What does this merge request do and why?

This MR implements a production-ready LRU in-memory cache for bind_tools() operations to resolve a performance bottleneck in the Duo Workflow Service.

The Problem

Google Cloud Profiler analysis revealed that bind_tools() operations consume 28.4ms and 6.03% of CPU time per request. This operation:

• Occurs 3-4 times per request (once per agent initialization)
• Runs synchronously in the asyncio event loop, blocking all other requests
• Has no caching, causing repeated expensive schema format conversions
• Limits theoretical maximum throughput to ~12 RPS

At 15 RPS load testing, the service experienced a 49% failure rate due to this bottleneck.

Root Cause: LangChain's bind_tools() performs expensive schema format conversion (OpenAI ↔️ Anthropic) on every agent initialization, with no built-in caching.

The Solution

This MR adds a thread-safe LRU cache that:

• Caches the result of bind_tools() operations by (model_id, tool_signature, tool_choice)
• Uses order-independent SHA256 hashing for stable cache keys
• Implements LRU eviction policy with configurable max size (default: 128 entries)
• Includes Prometheus metrics for monitoring (hits, misses, duration, size, evictions)
• Provides structured logging for debugging
• Is fully configurable via environment variables

Implementation Details

Files Created:

• ai_gateway/prompts/bind_tools_cache.py - Core LRU cache implementation

Files Modified:

• ai_gateway/prompts/base.py - Integration into Prompt class
• example.env - Configuration options (3 new env vars)

Related Issues

• Resolves: gitlab-org/gitlab#578158 - "Performance Bottleneck: Repeated pydantic.create_model() calls in bind_tools() blocking asyncio loop"
• Related: https://gitlab.com/gitlab-org/quality/quality-engineering/team-tasks/-/issues/3794 - "Agentic AI performance test execution and data gathering"

How to set up and validate locally

Numbered steps to set up and validate the change are strongly suggested.

Merge request checklist

• Tests added for new functionality. If not, please raise an issue to follow up.
• Documentation added/updated, if needed.
• If this change requires executor implementation: verified that issues/MRs exist for both Go executor and Node executor or confirmed that changes are backward-compatible and don't break existing executor functionality.