new gdb remote packet type

[Andrew Cagney]

This patch (if 'p' were implemented for gdbserver; I have this lying
around, as it happens) would make register fetches default to using
individual 'p' packets for every register; this would hurt latency, a
lot.
That isn't true.  The T packet should have previously returned all the 
important registers (and is needed anyway to make single step fast). 
This "p" would just fill in the gaps.

If after this we still have problems, we can investigate transfering 
registers in bigger chunks using qPart:<regset> (it was concluded that, 
for the moment, it is too bigger sledge hammer for this simple nut).

Robert, wouldn't it be good enough for you to work with
!reg->in_g_packet?
The original problem is that all registers are in the g-packet and that 
it was too big.

Andrew