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

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.