feat: [REL-10772] Implement Langchain provider for online evals

[edwinokonkwo]

Summary

This PR adds LangChain provider support to the LaunchDarkly AI Python SDK
The Python SDK follows a monorepo structure with packages split across:

• packages/core/ - Core AI SDK with client, models, tracking, and provider abstractions
• packages/langchain/ - LangChain specific provider implementation as a separate installable package
  This architecture allows users to install just what they need mirroring our TypeScript SDK's modular design.

Before:

python-server-sdk-ai/
└── ldai/
    ├── client.py
    ├── models.py
    ├── providers/
    │   ├── __init__.py
    │   ├── ai_provider.py
    │   └── langchain/          # langchain was bundled in
    └── ...

After:

python-server-sdk-ai/                  
├── packages/
│   ├── core/                          # Core SDK
│   │   ├── ldai/
│   │   ├── tests/
│   │   └── pyproject.toml
│   └── langchain/                     # LangChain provider is sperate package
│       ├── ldai/providers/langchain/
│       ├── tests/
│       └── pyproject.toml

[andrewklatzke]

Tracking the config ids that were pulled here would be a good facet to add imo

[andrewklatzke]

nit/suggestion: all_variables = variables or {}

[andrewklatzke]

I think this should be random.random() <= sampling_rate

A sampling rate of 0.1 should be 10% of the invocations, since the range of random() is 0-1, that leaves 90% of the results as greater than 0.1. So I think we want lte to capture 10%.

A sample rate of 1 should be 100%, so <= will always return true