HAMMER VFS - More cluster_read() fixes

HAMMER VFS - More cluster_read() fixes

* The recent cluster limiting code was improperly using raw block device
  offsets to calculate the largeblock mask used to prevent clustered I/Os
  from crossing a large-block boundary.

  raw block device offsets are not necessarily largeblock aligned, so
  this calculation failed to properly limit cluster_read()s and the
  result was an occassional clustered read would cross-over into an
  incompatible largeblock and later cause an overlapping buffer panic.

* Calculate the proper limit in the hammer_ondisk.c module and refactor
  hammer_io_read().

Reported-by: swildner