Don't do full drain on chunks containing future buffered data

[rzikm]

Description

This PR fixes bug where in certain situations, MsQuic would drop/corrupt data when received out of order + specific read timing from application side.

Testing

Failing test added to detect future regressions.

Documentation

N/A.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 85.33%. Comparing base (7bf265f) to head (bb006f4).
Report is 16 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4215      +/-   ##
==========================================
+ Coverage   84.22%   85.33%   +1.11%     
==========================================
  Files          56       56              
  Lines       15358    15359       +1     
==========================================
+ Hits        12935    13107     +172     
+ Misses       2423     2252     -171     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[rzikm]

We still have some failures on the stress tests even with this fix, (dotnet/runtime#100064 (comment)), let's not merge just yet, I will see if there is more to be done in this PR.

[nibanks]

This isn't quite the right fix. It should still call QuicRecvBufferFullDrain to indicate it drained everything that was read, but QuicRecvBufferFullDrain should be handling the fact that just because it read everything doesn't mean there isn't more after a gap to read.

[nibanks]

I am looking into what the right fix should be.

[rzikm]

I got more luck with this sort of change: rzikm@2beca9d

The gist of it is adjusting the ReadStart for multi/circular mode buffers, and moving the rest of the data for "single" mode buffers (not done in referenced commit, I just wanted to have something quic to run in stress env)

[nibanks]

We talked over IM and maybe using the partial drain is the right way to go. We'd just kind of redefine 'partial drain' here to indicate it's not partial of the read length, but of all (possibly discontiguous) data in the buffer.