The core problem here is how we prebuffer data that we read out of the cloudcache store. We do some advanced machinations with the underlying pipe where we keep our read pointer at the start, but push the examined pointer out to read in everything.

This works fine to stream all the data in, but causes a problem later if we end up doign multiple smaller reads over that as the small reads look like tehy're moving us backward versus the size of hte full read.

Note: this PR is a WIP. We are getting a new API in nerdbanks that will take care of this entire scenario of prebuffering safely. However, this PR at elast shows a test for this case, as well as a fix that works for us (albeit with an unfortunate non-pooling of pipe backing memory).