This reverts commit b4c738eb.

The race condition didn't trigger when I tested it, so I assumed I had
solved it...

Some recipes require the headers, but they don't explicitly say
so. Parallel make (-j`nproc`) might start compiling the libs when the
headers aren't done yet.

Add sys_timeb header

- Remove incorrect header styles
- Use export.replace where header style was previously needed
- Check compilation of tests using system gcc

We cannot use ?= because CC is set by default to "cc". Therefore
CC was never set. In addition we need to export the variable
in order to have the rust cc crate pick it up. Otherwise it is
only used by openlibm where it is explicitly passed.

These prerequisites are GNU Make terminology.

This change forces the libc rule to be executed first so that the headers that
the libm rule needs are available. The original setup was using 'normal'
prerequisites which occasionally resulted in bizarre build breakage,
especially on multi-core build hosts. Seen often on my 8-way SMP build
host.

Note that this doesn't impede parallelisation of each rule indepent of
the other. It just serializes the rules themselves.

This fixes: redox-os/relibc#127