Sharrow updates

[jpn--]

• Add dask_scheduler argument for loading data into shared memory. Previously loads were always done in parallel using the threaded scheduler, but now the user can choose a single-threaded synchronous option instead.
• Optimize equality and inequality comparisons on categorical data where the comparator values are not in the categories. These cases will be universally True or False, so there's no need to encode and compile an actual comparison.