Update relative offset resolution to be saturating

[nicholastmosher]

Right now, using relative offsets Offset::from_beginning(X) or Offset::from_end(X) when fetching from a partition have undesirable behavior when X is greater than the number of records in that partition. When this happens, a stream request will just hang. Instead, I think that the relative offset resolution should be saturating, meaning that:

If n is the size of the partition, then

• When we fetch using Offset::from_end(X) and X > n, then the fetch should resolve to absolute offset 0
• When we fetch using Offset::from_beginning(X) and X > n, then the fetch should resolve to absolute offset X-1

The end result of this is that no call to fetch should ever cause the client to hang.

Related: #1007

[sehz]

find root cause