depth.ocean_floor is very memory hungry

[mx-moth]

Need to profile this properly, but from casual inspection doing depth.ocean_floor() will allocate much more memory than it actually needs.

Some significant drops in memory usage have been observed by splitting the dataset along the time axis, flooring each slice one at a time, then merging it back together.

Further benchmarking is required to really check out what is going on. Some robust tests and benchmarking is required to validate whether slicing along time is broadly speaking an improvement, and whether it works in all cases.