Add IR based JIT

[dstogov]

This adds JIT for your VM bytecode based on https://github.com/dstogov/ir
I just liked to check how IR framework does its work. The first naive version was 10 times slower than "native" code, but the obvious stack resolution made JIT-ed code almost as fast as "native" (~2-2.5% slower). The SSA construction removed even this gap.

It shouldn't be a big problem to port this to LLVM. SSA construction should be done by "mem2reg" pass on top of "jitted_ir_var".

Also, I found that a weak-typed PHP interpreter is significantly faster than your best asm optimized version.
So it's definitely possible to write much faster interpreters...

@atakua you may be interested.

Benchmark Results

native                                   0.601
asmopt                                   2.986
predecoded                              21.943
subroutined                             15.425
switched                                24.787
tailrecursive                           14.670
threaded                                25.444
threaded-cached                         18.190
translated                              20.212

php -d opcache.enable=0                  2.188
php -d opcache.jit=0                     1.968
php -d opcache.jit=function              1.005
php -d opcache.jit=tracing               1.011

jited_ir                                 6.277
jited_ir_stack                           0.616
jited_ir_var                             0.613
jited_ir_ssa                             0.601