Bump parquet and use arrow-js-ffi

[kylebarron]

arrow-js-ffi is a much more memory-efficient option for moving Arrow data from Rust to JS because it doesn't require a full buffer copy on the Wasm side to construct an Arrow IPC buffer.

This is pretty much blocked on apache/arrow-rs#5356 or some workaround to it. Because of the way deck.gl handles GPU-based picking, it can have a maximum of 256 layers because their indices are packed into a single channel. We take great care on the Python side to choose a record batch size that limits the number of pickable layers to a max of 256. But because arrow-rs does its own rechunking, we're left with tons of tiny record batches of 1024 rows each, which blows up deck.gl (this is from the parquet-wasm bump, not related to using arrow-js-ffi).

Also blocked on fixing parquet package exports (ref kylebarron/parquet-wasm#414)

[kylebarron]

Note that we only need a param for the record batch size. In Lonboard we already send separate Parquet files that each contain a single logical Arrow batch. So we really just want to set the record batch size to a very large number.

[kylebarron]

I published parquet-wasm 0.6.0 and it would be good to test this again. We should bump parquet-wasm regardless of whether we enable arrow-js-ffi.

[kylebarron]

Superseded by #477