Lower initial dict/set size from 64->8

[kmod]

llvm's DenseMap/DenseSet have a hardcoded minimum size of 64 buckets, which I think ends up wasting a lot of memory for us. They also take longer to iterate over (more empty buckets to skip), which is probably not as big a deal in the macrobenchmarks but shows up in some microbenchmarks.

So, fork their implementation and lower the minimum to 8 (parameterizable now). Surprisingly effective:

           django_template3.py             2.8s (2)             2.8s (2)  -0.8%
                 pyxl_bench.py             2.3s (2)             2.3s (2)  -0.1%
     sqlalchemy_imperative2.py             2.8s (2)             2.7s (2)  -4.4%
       django_template3_10x.py            17.5s (2)            17.1s (2)  -2.6%
             pyxl_bench_10x.py            17.5s (2)            17.4s (2)  -0.4%
 sqlalchemy_imperative2_10x.py            21.0s (2)            19.7s (2)  -6.2%
                       geomean                 7.0s                 6.8s  -2.5%