switch to parquet to lower memory footprint

[cornhundred]

Currently we are loading data from a CSV and we have to load the entire CSV into memory. Since this example is running on mybinder, which has limited memory ~1GB (voila-dashboards/voila#521 (comment)), we want to be careful with memory usage.

Switching to parquet (and adding the pyarrow requirement) will enable us to load data from select cells (or genes in a transposed parquet) instead of loading everything.