Bug 358

Our CDR stream classes read a complete message from the network and then process
the message in memory. For some applications it could be more efficient to read
enough of the message as to start demarshaling and then simply:
+ Copy the rest of the message directly to another socket, minimizing latency in
gateways (this was suggested by James Megquier from BBN)
+ Drop the message because it is not important (ex: late replies to timed out
request.
This shouldn't be too hard to implement: a backing store class will hide the
source of the buffer and read it on demand, much like iostreams work; in the
simple case the backin store will read from a fixed size buffer (though a chain
of buffers is also interesting); in the socket case the backin store could read
using the Transport class from the pluggable protocols framework.
Writing to a backing store is harder because the message size is on the first
few bytes, and the real message size is not known until it is completed. Using
an iterator to record the placeholder left for the message size and some kind of
upcall to let the application write the total message size before flushing
*could* solve the problem, but that requires more thinking.