Add -mwin32 flag on Windows targets

[abonander]

See time-rs/time#45.

I figured adding it to the defaults would be the most controllable way of implementing this fix. That way users can opt-out if it's inducing unwanted behavior. I'll reboot into Windows in a second to test this.

Edit: I can't exactly attest to the fix as Cargo silences the prints from gcc-rs, however it does build correctly even after removing the patch I made to time, so I presume it works.

[alexcrichton]

This may actually prefer to go in the compile_library function because otherwise this may get overridden by accident. This should also use the value of the TARGET environment variable instead of cfg! due to cross compilations that may be happening.

[abonander]

What value of TARGET should I match then?

[alexcrichton]

target.contains("windows") should probably be good enough for now.

[abonander]

@alexcrichton Updated. r?

Couple notes:

• Config currently doesn't need an explicit impl for Default since it can simply derive it (default for Vec<T> of any T is an empty vector).
• All the .as_sllice() calls are redundant because of auto-deref.
• Is calling .to_string() on a string slice better than ToOwned::to_owned()? The former has to go through the formatting framework while the latter just copies into a new allocation. Not sure how it optimizes down.

[alexcrichton]

Config currently doesn't need an explicit impl for Default since it can simply derive it

I'd be down with that!

All the .as_sllice() calls are redundant because of auto-deref.

Ah, the wonders of Rust code written a few months ago :)

Is calling .to_string() on a string slice better than ToOwned::to_owned()?

For now on very large strings, it will be slower. Once rust-lang/rfcs#526 is implemented, however, this will largely no longer be true (the performance won't be the same, but the gap will be of a constant factor, not O(n))