The last condition of this test no longer applies because Git will
repack files in a different way than it used to.

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>

We will start hashing loose objects downloaded from the remotes, but we
need to include a package (SharpZipLib) that can handle the headers
included in the ZLib compression used by Git. .NET's default
decompression algorithm cannot handle those optional headers,
unfortunately.

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>

After we write the data to disk, check the file that it matches the
correct hash before moving it into place. If the hash is wrong, then
include details about where the data is stored for reference.

This can be marked as optional to make existing unit tests work, but the
logic is covered by a new set of unit tests.

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>

The loose object download needs to verify the objects being sent across
the wire before moving them to their "real" locations in the object
directory. The previous implementation wrote them to disk then read from
that file to hash the contents.

Instead, create a new "SideChannelStream" that injects between the
responseStream and the inflater, but sends all of the data that is read
out to the filestream. Here is a diagram of the stream sequence:

  responseStream -> sideChannelStream -> inflateStream -> hashingStream
                           \                                   |
			    \                                 \|/
			     tempFileStream                devNull

By copying all data from the hashingStream into devNull, we ensure that
the responseStream is drained and written to the tempFileStream while
simultaneously inflating and hashing the object contents. We then
amortize this computation during the window that we are downloading data
over the network, minimizing the potential performance impact.

Helped-by: Matthew Cheetham <mjcheetham@github.com>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>