feat: use media url directly in both LMM and code sandbox

[yzld2002]

We used to always use the local media:

• encode it as base64 and talk with OpenAI
• upload it to the code interpreter before execution

This PR adds the support to directly use the media URL

1. OpenAI supports both base64 and URL format of the image
2. Added the support to download from URL in the tool load_image and extract_frames. Although the moviepy can directly use the URL, but after testing, looks like download it first seems more stable.
3. During write_plan, we need to feed the LLM one frame of the video, we assume there is a same url that only differs in the suffix. (https://xxxxx.mp4 -> https://xxxxx.png)

Local run:

• image

• video

[dillonalaird]

LGTM will slack you some extra details on this