Rework the file handling here to explicitly probe the type of input

Rework the file handling here to explicitly probe the type of input
we're using and use that to determine an explicit I/O strategy.  This
was largely inspired by an email exchange with Duane Hesser, who
clarified some of the issues involved in doing high-quality tape
handling.  I think the approach here will make it much easier to
provide optimized I/O strategies for tape and sockets.

Because of these changes, reading the directory of an ISO image stored
on a raw device (via "tar tvf /dev/cd0", for example) is about 100x
faster due to a combination of better detection of "disk-like" devices
and a more suitable strategy for handling forward skip requests.
Extracting tar devices stored on one disk drive onto a physically
separate drive should also be significantly faster because we
now do block-size cheating on disk-like devices.

SVN-Revision: 1948