implement first version of vlm predicate classifier

[lf-zhao]

Start a PR on using VLM for predicate evaluation, replacing original classifier for e.g., On and Inside. Working on next steps, but create one PR for keeping manageable size.

• use rich package to make structured logging
• add openai VLM interface
• add some test for it
• add fields in State to store current images, objects
• add some logic to save these info to state
• implement VLM predicate classifier with the interface
• run and test on LIS Spot!