Improve performance of memory size partitioning

[smcguire-cmu]

Currently, partitioning with a memory size bytes threshold is quite slow. After discussion with Olivia, there were a few ideas of different potential ways to improve this performance including:

• For fixed length columns (int, float, etc.) use known size instead of computing per row
• Use vectorized methods to get size/length of dynamic length columns if they exist
• Do size calculations per histogram pixel instead of per row.

We should try to benchmark and improve performance with any of these methods if they work, or any other ideas.