Speed up wal write

Right now we do 4 memcpy and 4 syscalls to make sure each user request which changes data reaches the disk.
Reduce this overhead by using shared memory and producer-consumer pattern, rather than the message-passing paradigm
used today.
Maintain "the last written lsn" across the front-end and the back-end, and alert all fibers waiting on wal-writer
based on the lsn they are waiting for (i.e. do not maintain an individual inbox in each fiber for that purpose).

Use offset_t as lsn, not consequetive lsn id -- that will simplify lsn generation.