Merge davem's arith improvements

[rurban]

It's not better than my faster overflow checks, but it's at least something in an orthogonal area. It tries to avoid the overflow checks at all, earlier.
The 4 commits starting with 20e9643

• 20e9643 split pp_postdec() from pp_postinc() and improve
• 4c2c312 split pp_predec() from pp_preinc() and improve
• 230ee21 faster add, subtract, multiply
• edba15b make SETi/u/n, (X)PUSHi/u/n more efficient

Need to benchmark that if it's not actually slower than fast overflows at all. It adds a lot of branch code before that. If it's slower use that only with older C compilers without HAS_BUILTIN_ARITH_OVERFLOW

The old fastarith branch from February started here: 8b6139a

[rurban]

The benchmarks confirm my suspicion:
Before: 227 sec, After: 336 sec, 48% slower
So this patch is only useful for older compilers without overflow builtins

tested on darwin x86_64 with gcc-5.2