Rewrite the driver to handle 64kb transfers

Rewrite the driver to handle 64kb transfers

Although several codepaths in the kernel such as coredump
and buffercache read-ahead feature assume that underlying
hardware is capable of handling 64kb transfers without any
issues, xbf was setup to rely on a single descriptor per
transfer which limited the maximum size of an individual
transfer to 11 4k segments amounting to 44k bytes.

To avoid overbooking, a metadata object is allocated for
each transfer to keep track of associated descriptors
limiting the maximum amount of outstanding transfers to
half the ring size.

The issue was reported by Dan Cross <crossd at gmail.com>,
thanks!