apr_file_t is an abstraction that covers the whole range of what can be
a file (including pipes etc.) and how it can be used (multi-threading,
overlapped I/O etc.).  This limits the efficiency of its implementation.
 
Design goals for svn_file_t:

* Build upon non-buffered apr_file_t to get the platform support without
  any of the buffer management overhead.
* Use it for "real files" with at most one writer process only.  This
  allows for efficient file length and EOF detection.
* Map an unlimited number of instances to a limited number of file handles.
  Handles to same file may be shared (depending on flags) and non-locking
  file handles may be closed temporarily.

* Use (up to) two buffers instead of one per file to allow for efficient
  traversal in both directions.
* Internalize seek() operations, i.e. don't pass them on to the OS.

The current code is completely untested and does not fully comply with
the design goals listed above.