ENH: Add allocation boundaries for inline task memory requirements

[seberg]

gh-11 applies the quick fix to work with the new one-pool implementation of legate. However, not providing any upper bound for the memory allocation needs prevents parallel execution of other tasks.

While it is a bit hard to know if libcudf is doing temporary allocations, at least for tasks such as unary/binary operations this should be easy.
For a task like CSV reading or join/aggregates it is harder since the (current) worst case scenarios for memory use can be surprisingly high, although, maybe even then it is useful to add the estimate. E.g. when reading a small CSV file.