Fixed semantics of CLUSTER SETSLOT, SELECT now only denied in cluster mode if selected DB is not 0 so that MIGRATE still works well.

implemented a different approach to IO scheduling, so object->storage is no longer used, instead there is a queue and hash table of IO tasks to process, and it is always possible to know what are the scheduled and acrtive IO operations against every single key.

touched key for WATCH refactored into a more general thing that can be used also for the cache system. Some more changes towards diskstore working.

Do not update the LRU info on key lookup when we have a saving child. With this trivial change the additional memory used while saving with a background child in presence of many read operations is zero.

Don't increment dirty on expireIfNeeded() as natural expires are not considered database changes. This will avoid useless read-only commands in the AOF file as a result of, for instance, a GET operation triggering an expirIfNeeded() call resulting in an expired key removed.

Object approximated LRU algorithm enhanced / fixed / refactored. This is used for the VM currently but will soon be used for maxmemory expiring.

fixed an alignment problem with time_t is 32 bit, long is 64 bit, and arch is sparc or any other where unaligned accesses will result to sigbus

Fix re-enabled again, I forgot to check if VM was enabled before calling handleClientsBlockedOnSwappedKey()

Fixed a race condition in VM happening when a key was deleted while there was a client waiting for this key to be resumed from swap to memory. The client would hang forever.