MPI+SparseTimeFunction: More efficient _dist_gather

[FabioLuporini]

In the context of checkpointing, it's a significant overhead that upon returning from C-land we redistribute the entire SparseTimeFunction while potentially only a relatively small number of time iterations have been computed

This should be easily fixable by plumbing the args down to _arg_apply and then to _dist_gather so that we only retain the written region of the data array

[tjb900]

Probably a very similar fix if so, but does this potentially apply to _dist_scatter before going to C also?

[mloubout]

Should we drop the "gather" completely and only scatter once and be done?

[FabioLuporini]

I think there might be side effects if one expects certain data to be on a certain rank while it actually is somewhere else.

Perhaps we can:

• scatter upon startup (ie very first SparseFunction initialization)
• gather on-the-fly each time sf.data is accessed (ie we return a view)

this way would be basically 0 overhead during forward and backward propagation