Improve byte slice encoding

[frankmcsherry]

This PR introduces a new encoding of a sequence of byte slices, intended to provide more efficient access to the slices when one does not plan on examining all of them. Specifically, to iterate over the byte slices one only has to touch a compact prefix of the memory, in which we record the offsets of each of the slices into the larger memory. This contrasts with the current encoding: a sequence of length delimited byte sequences, which means one must traverse them to find any one slice, or to count their number.

This is a breaking change, as the encoding only works for A: AsBytes, rather than an iterator over slices. We could extend it to an iterator that supports clone(), but we need to perform three passes to extract the information (number of offsets, values of the offsets, and the bytes themselves).

More testing and integration is needed before this is published.

[frankmcsherry]

One integration observation is that timely has hard-coded the prior serialization approach, in part because without #28 there is no other way to write into a W: Write.