Further optimize code size

[Amanieu]

I had a look at #276 and thought I could do better!

There basically two areas where an improvement can be found:

• LLVM generates pretty poor code for the default implementations of PartialOrd::{lt,le,gt,ge} which forward to partial_cmp. So I added versions which perform the desired comparison directly.
• The target and location are almost always compile-time constants and can be passed in as a single argument which points to a static constant. This unfortunately doesn't work for the log level since we first need to make a local copy of it so that it is only evaluated once.

A main function containing nothing but warn!("hello world") shrinks from 160 bytes to 98 bytes in x86_64 with this change.

[sfackler]

Thanks!

[alexcrichton]

@sfackler mind doing a release with this in it? It may actually be a bit of a perf boost as well for the comparisons!

[sfackler]

Done!