The storage strategy per-object is the same as NORMAL, but instead of having
immutable chains of hidden classes so that we can share them between instances,
a SINGLETON hcls is unique to the single object that it is for.  This means
that it can be mutable.

This has the advantage that we don't have to track all the parents
of the hidden class.  This is especially helpful since the cases that
it's used for (Python classes and modules) tend to have a large number
of attributes (>100 not uncommon).

We could also use the DICT_BACKED strategy for those cases, but compared to
DICT_BACKED, SINGLETON should have much better lookup performance since
it is patchpoint-compatible.  Adding new attributes is probably slower,
and deleting attributes is probably much slower, but I think those are
hopefully rare.

This cuts non-jit 'import pip' memory by 15%, from 113MB to 98MB.