llvm tier: reduce scratch size

[undingen]

With the new frame introspection we don't need to save additional space for spilling vars which live in regs (except for deopts).
And for deopts we will never need more scratch than the one required for spilling the args.

                                   upstream/master:     origin/pp_stack:
           django_template3.py             2.1s (2)             2.1s (2)  -0.4%
                 pyxl_bench.py             1.8s (2)             1.8s (2)  +0.2%
     sqlalchemy_imperative2.py             2.2s (2)             2.2s (2)  -0.4%
                pyxl_bench2.py             1.1s (2)             1.1s (2)  -1.1%
       django_template3_10x.py            14.0s (2)            13.9s (2)  -0.6%
             pyxl_bench_10x.py            13.7s (2)            13.6s (2)  -1.0%
 sqlalchemy_imperative2_10x.py            16.9s (2)            17.0s (2)  +0.3%
            pyxl_bench2_10x.py             9.3s (2)             9.2s (2)  -0.8%
                       geomean                 4.8s                 4.8s  -0.5%

[kmod]

Seems reasonable; not sure if anything related to this has changed so that we would want to re-test it.

[undingen]

I checked again perf numbers are:

                                   upstream/master:     origin/pp_stack:
           django_template3.py             2.1s (4)             2.1s (4)  -0.1%
                 pyxl_bench.py             1.8s (4)             1.8s (4)  +0.9%
     sqlalchemy_imperative2.py             2.2s (4)             2.2s (4)  -0.4%
                pyxl_bench2.py             1.1s (4)             1.0s (4)  -1.2%
       django_template3_10x.py            13.8s (4)            13.8s (4)  +0.2%
             pyxl_bench_10x.py            13.3s (4)            13.2s (4)  -1.2%
 sqlalchemy_imperative2_10x.py            17.0s (4)            17.0s (4)  -0.2%
            pyxl_bench2_10x.py             9.1s (4)             8.8s (4)  -2.4%
                       geomean                 4.7s                 4.7s  -0.5%