Major new memcached goodie checked into svn

[Steven Grimm]

The "multithreaded" branch in svn now has, as the name implies, a 
version of memcached that can be compiled as a multithreaded server. We 
are running memcached on 4-core servers; whereas before we had to split 
memcached up into separate smaller processes when we bottlenecked on 
CPU, now we can run one large process and take advantage of increased 
multi-key "get" batching, etc.

I made the change such that you can continue to compile memcached as a 
single-threaded server if you are not bottlenecking on CPU time; in 
fact, single-threaded mode is the default, since it is sufficient for 
almost all memcached installations and runs with less overhead (no need 
to synchronize between threads.) That was a major design goal since it 
will, if people are satisfied with the changes, allow this version to be 
folded back into the memcached trunk at some point without breaking 
platforms that don't have thread support.

An overview of the change is in doc/threads.txt. Briefly, the change 
involves assigning each incoming connection to one of a small fixed 
number of worker threads; each worker thread acts like a miniature 
memcached instance, using libevent to divide its time among all the 
clients it's in charge of. The UDP socket is shared among all the worker 
threads.

For now this is perhaps best considered an experimental change, though 
we are running it successfully on a number of our production servers.

If other people could give it a whirl in either multi- or 
single-threaded mode, that'd be great. I did most of the development on 
Mac OS X, and the production servers are running Linux on AMD64, but it 
is essentially untested on platforms other than those. In theory it 
should work fine on any platform that supports both pthreads and 
libevent, but it'd be nice to have some real-life success reports!

-Steve