Lightweight Legion

[manopapad]

Opening this on behalf of @lightsighter. The goal is to disable parts of the runtime that are irrelevant for the current configuration, so we don't pay for those at all:

cpu-only -> no dependence analysis and no coherence analysis
1 GPU -> no dependence analysis, yes coherence analysis but a single equivalence set
N GPU (single node) -> dependence analysis and limited coherence analysis
multi-node -> back to normal

[lightsighter]

Making a note that I am less convinced that this is necessary after this analysis. We will discuss in the Legate engineering meeting.

[lightsighter]

Some evidence suggests that Legion is fast enough as-is for single-CPU/GPU use cases. For single-process multi-GPU cases which lightweight Legion could also help with, we need to have confirmation that users actually want to run in that mode. @manopapad to decide if that is actually useful or if it doesn't matter because most use cases will be running in a GPU-per-process configuration (in which case Lightweight Legion would not be of any help).