Implement a GroupTotal trait and operators

[frankmcsherry]

Issue #74 calls for specialized implementations for totally ordered timestamps. The group and group_u operators are currently very complex, in part due to the complexity of partially ordered times. This is a non-trivial implementation (much more complex than CountTotal or DistinctTotal), but there are at least a few important simplifications we can exploit:

1. The times at which the group variants must be re-evaluated are exactly those found in the input differences. This is unlike for partially ordered times, which have a more complicated logic to determine further synthetic times that may produce output differences.

2. The running accumulations for each key need not (i) track the meet of subsequent times (it would be just the min, equal to the next time in sorted order), nor (ii) maintain an intermediate collection of updates that have been advanced up to this meet (it can instead maintain a simpler running accumulation of the collection in question).

There are still a large number of subtleties, and the current group.rs file may not be the best starting point due to its complexity (both intrinsic complexity, and questionable structure). I recommend if anyone wants to try and work on this issue we should get in touch and map out the right way to spec out how it should work.