This works around strict aliasing rules letting some versions of
GCC (particularly on RHEL 6) thinking that they can skip updating the
size of the array when calculating the next element's offset.

Preallocating two commits doesn't make much sense as leaving allocation
to the first array usage will allocate a sensible size with room for
growth.

This preallocation has also been hiding issues with strict aliasing in
the tests, as we have fairly simple histories and never trigger the
growth.

We used to assume a refspec would only have an asterisk in the middle of
their respective pattern. This has not been a valid assumption for some
time now with git.

Instead of assuming where the asterisk is going to be, change the logic
to treat each pattern as having two halves with a replacement bit in the
middle, where the asterisk is.

When transforming a non-pattern refspec, we simply need to copy over the
opposite string. Move that logic up to the wrapper so we can assume a
pattern refspec in the transformation function.

In order to connect to a remote server, we need to provide a path to the
repository we're interested in. Consider the lack of path in the url an
error.

When the stream writing function was written, it assume that
libssh2_channel_write() would always write all of the data to the
wire. This is only true for the first 32k of data, which it tries to
fit into one ssh packet.

Since it can perform short writes, call it in a loop like we do for
send(), advancing the buffer offset.

To make sure that items returned from pool allocations are aligned
on nice boundaries, this rounds up all pool allocation sizes to a
multiple of 8.  This adds a small amount of overhead to each item.

The rounding up could be made optional with an extra parameter to
the pool initialization that turned on rounding only for pools
where item alignment actually matters, but I think for the extra
code and complexity that would be involved, that it makes sense
just to burn a little bit of extra memory and enable this all the
time.