Couple fixes for the windows backend

[fgenesis]

First time dmon user, and it immediately blew up! (When trying to set a watch to a directory that didn't exist. win10.)
So here's a bunch of fixes to make it work more reliably.

Copied from the commit:

• _dmon.modify_watches is gone. Its use introduced more hazard than gain.
• Fix possible race conditions when watch creation failed
• Actually handle errors properly instead of corrupting internal state
• Don't keep the global mutex held while lazing around in Sleep()
• Since watches are always deleted in unwatch, there's no use trying to keep a pool of allocated watches around for re-use

I've been thinking to remove the weird "freelist" construct too, if you're ok with that.

Because we can just loop over the (compile-time-bounded) array of active watches and pick the first index that's NULL. Since adding and removing watches is nothing done on a per-frame basis a couple loop iterations don't matter. I'd prefer the gained simplicity.

_dmon.quit should be at least volatile (done now), but an atomic would be even better. Comments?

Lasty, do you mind if i get rid of the #include <stdint.h> for the windows backend? Because I like backwards compatbility with older compilers and this is the single most annoying header to break MSVC <= 2012-ish compatibility.

I'll add some extra commits to the PR when i get the go-ahead from you.

[septag]

Thanks!

[mwestphal]

_dmon.quit should be at least volatile (done now), but an atomic would be even better. Comments?

FYI I tried to make it atomic but it then refused to compile. I did not investigate more.

[septag]

I'll try to make all those atomics soon, haven't got the time yet

[septag]

I removed all volatile vars and now using atomics