DEBUG DIGEST implemented, in order to improve the ability to test persistence and replication consistency

very strong speedup in saving time performance when there are many integers in the dataset. Instead of decoding the object before to pass them to the rdbSaveObject layer we check asap if the object is integer encoded and can be written on disk as an integer.

added overflow check in the double -> long long conversion trick to avoid integer overflows. I think this was not needed in practical terms, but it is safer

If a float can be casted to a long long without rounding loss, we can use the integer conversion function to write the score on disk. This is a seriuous speedup

fixed a problem leading to crashes, as keys can't be currently specially encoded, so we can't encode integers at object loading time... For now this can be fixed passing a few flags, or later can be fixed allowing encoded keys as well

long long to string conversion speedup applied in other places as well. Still the code has bugs, fixing right now...

added specialized function to compare string objects for perfect match that is optimized for this task

while loading the rdb file don't add the key to the dictionary at all if it's already expired, instead of removing it just after the insertion.

swap objects out directly while loading an RDB file if we detect we can't stay in the vm max memory limits anyway

add sanity check to zunionInterBlockClientOnSwappedKeys, as the number of keys used is provided as argument to the function

make prototype of custom function to preload keys from the vm match the prototype of waitForMultipleSwappedKeys

AOF writes are now accumulated into a buffer and flushed into disk just before re-entering the event loop. A lot less writes but still this guarantees that AOF is written before the client gets a positive reply about a write operation, as no reply is trasnmitted before re-entering into the event loop.

AOF is now rewritten on slave after SYNC with master. Thanks to @_km for finding this bug and any others'

Pub/Sub API change: now messages received via pattern matching have a different message type and an additional field representing the original pattern the message matched