Support the Arrow-based data transport.

[whitphx]

• st.experimental_data_editor introduced with 1.19.0 relies on PyArrow.
• The legacy transport may no longer be supported in the future.

[whitphx]

st.dataframe is being improved so much with Arrow-based data communication.

[lukasmasuch]

Some ideas from a private discussion on this topic:

I think what would already help is any way on how we could get the pandas dataframe serialized to the IPC/feather format (bytes) on Python side. We don't really need any other functionalities from Pyarrow. One pyodide supported library that might get us on step closer is fastparquet. This would probably allow us to serialize the dataframe into the parquet format. And maybe we can use something like parquet-wasm (or arrow-wasm) to actually convert it from parquet into the IPC format on JS side. Once we have this in IPC format, we can load this in our frontend (here) and everything should work fine

What we currently do in Streamlit is:

1. Backend: Convert any data to Pandas DF
2. Backend: Convert Pandas to Arrow Table
3. Backend: Serialize Arrow Table to bytes (IPC format)
4. Frontend: Load bytes via tableFromIPC

For stlite, we might be able to adapt this process so that:

1. Backend: Convert any data to Pandas DF
2. Backend - stlite-specific: Serialize Pandas to parquet file (via fastparquet)
3. Frontend - stlite-specific:: Load parquet file and convert to IPC bytes via parquet-wasm
4. Frontend: Load bytes via tableFromIPC

[whitphx]

@lukasmasuch Hi, thank you very much for such a detailed comprehensive guide of migration!
This is the changes I made on the Streamlit codebase forked for stlite: whitphx/streamlit#3. Can you please take a look on it and see if it makes sense?
Also, you can see st.data_frame and st.data_editor are working in the preview envs deployed in this PR: #601