Fix #86: save fpu state on entry to alloc
With gcc, alloc() is compiled to use xmm registers before we can save them. So, let's save the xmm registers in alloc_overflow_sse2 which calls alloc. The issue we see with gcc not working is gone now.
While we're at it, use the
__attribute__ option to get a 16-byte
aligned area where we can save the fpu state.
And also set optimization to
-O2 for linux effectively undoing #68 (closed).