This was causing modelsim to freak out.
It had no effect on hardware because it was a don't care situation.
The new code makes explicit that we don't care about the underflow value.

Long-term, once the TX buffering is fixed, this will not be a problem.
In the meantime, it still needs to work. This re-hack fixes the hack.

I have been convinced by rubini and the linux kernel sources.
Opaque types for pointers to struct just make code confusing.

add blocking interface via magical "eb_block" callback
  => causes close to block until complete and return status
  make blocking the default for C++ API
add format pretty-printer helper functions:
  eb_width_data, eb_width_address, eb_format_data, eb_format_endian
add matching parser functions:
  eb_width_parse_address, eb_width_parse_data
use long for delays (int is too short on some platforms)
remove eb_device_flush
  => for now hack around lack of write flow control
add convenience methods eb_device_{read,write}