This is an attempt at improving allocation of str (PyUnicode) objects.
For that it does two things:
1. make str objects PyVarObjects, so that the object is allocated in one
pass rather than two
2. maintain an array of freelists, each for a given str length, so that
many str allocations can bypass actual memory allocation calls

There is a ~10% speedup in stringbench, and a slight improvement in
pybench (as invoked with "./python Tools/pybench/pybench.py -t String").
Also, memory consumption is a bit lower when running those benchmarks.