feat: Proxied Model Support

[JGalego]

This PR adds support for models served behind OpenAI-compatible proxies e.g. LiteLLM or Bedrock Access Gateway (missing image generation support). The constraint is to change as little as possible of the original implementation.

Example (LiteLLM >> Amazon Bedrock)

1. Set up AWS credentials

   export AWS_ACCESS_KEY_ID=...
   export AWS_SECRET_ACCESS_KEY=...
   export AWS_SESSION_TOKEN=...

2. Start LiteLLM proxy

   litellm --config config.yaml

   # config.yaml
   # Adapted from https://litellm.vercel.app/docs/proxy/configs
   
   model_list:
     - model_name: claude3-sonnet
       litellm_params: 
         model: bedrock/anthropic.claude-3-sonnet-20240229-v1:0
         aws_region_name: us-east-1
         temperature: 0.0
     - model_name: sdxl
       litellm_params: 
         model: bedrock/stability.stable-diffusion-xl-v1
         aws_region_name: us-east-1
   
   litellm_settings:
     drop_params: True
     modify_params: True  # hack to get around Claude's Message API restrictions
   
   general_settings: 
     master_key: correct-horse-battery-staple

3. Set up OpenAI and ChatDev env vars

   # OpenAI
   export OPENAI_BASE_URL=http://0.0.0.0:4000
   export OPENAI_API_KEY=correct-horse-battery-staple
   
   # ChatDev
   export CHATDEV_CUSTOM_MODEL=claude3-sonnet
   export CHATDEV_NUM_MAX_TOKEN=200000
   export CHATDEV_CUSTOM_IMAGE_MODEL=sdxl

4. Run ChatDev

   python3 run.py --task "Design a simple game of tic-tac-toe" --name "TicTacToe"  --org "THUNLP" --config "Default"

chatdev_bedrock.mp4

[JGalego]

Hey @thinkwee, reaching out b/c you're currently listed as the top contributor.
Can we have someone review this PR?