Previously, render parts were stored in a linked list and every tile update
searched the entire list for the correct part. As a result, the overhead
of searching tiles increased quadratically w.r.t. the number of tiles.

By hashing the parts based on their location, this operation is much faster,
significantly reducing the tile update overhead for small tiles and/or large
renders.

For example, rendering an empty scene in 1080p at 1spp and 8x8 tiles goes
down from 9.22sec to 1.45sec on my laptop.