major bjit improvements

[undingen]

First patch of baseline JIT improvements.
the major changes are:

• it reenables the optimization that we don't use the vregs for compiler generated vars which don't get used after the current basic block
• allows usage of r12 and r15 in the bjit
• jit bjit code into 31bit address space

This makes the generated code look much more reasonable and is a good foundation for further improvements.

I have a second tier of patches which I will put into another PR which makes the bjit reuse some of the vregs which improves the performance further.

                                   upstream/master:    origin/bjit_opt2:
           django_template3.py             2.8s (4)             2.6s (4)  -6.5%
                 pyxl_bench.py             2.8s (4)             2.4s (4)  -13.7%
     sqlalchemy_imperative2.py             2.7s (4)             2.7s (4)  -2.7%
                pyxl_bench2.py             1.2s (4)             1.2s (4)  -2.9%
       django_template3_10x.py            13.8s (4)            13.5s (4)  -2.5%
             pyxl_bench_10x.py            18.3s (4)            18.1s (4)  -0.9%
 sqlalchemy_imperative2_10x.py            19.3s (4)            19.3s (4)  +0.0%
            pyxl_bench2_10x.py            10.7s (4)            10.2s (4)  -4.5%
                       geomean                 5.8s                 5.6s  -4.3%

some of the minibenchmarks in -SI mode:
raytrace.py        21.29s -> 15.22s
django_template.py  3.75s ->  3.24s
richards.py         4.57s ->  3.20s

[undingen]

should be ready for review

[kmod]

lgtm too :) but let's run it through the CI as well