Explore using virtual Zarr approaches to improve Eva performance

[ashiklom]

We may be able to get significant performance gains by creating "virtual Zarr" stores (e.g., Icechunk virtual dataset) from NetCDF files (especially, when these are big stacks of NetCDF files). We should experiment with (1) what an Eva driver for virtual datasets might look like from an implementation standpoint; and (2) how much performance we pick up from doing this.

Assigning @brandonzhu09 to look into this.