Changed the BIGMSGS semantics.

We now identify the memory location of the start of message body
during serialization, so we don't have to measure it twice.  Then
we store the body in the BIGMSGS table only if the total message
size would exceed 4000 bytes.  This assumes that memory and disk
pages are 4096 bytes, so we want to optimize for a single I/O
operation in most cases.