This adds seven benchmarks for GC paths that are not well covered by the existing suite:

• gc_collect_large uses 10,000 independent cycles, while gc_collect_self_lists uses a minimal synthetic graph to expose per-object collection overhead.
• gc_weakref_recheck, gc_resurrection_partial, and gc_resurrection_full cover rechecks after a weakref callback, partial resurrection, and full resurrection.
• btree_gc_only_frozen covers full collection with a frozen startup heap.
• gc_tuple_survivors times collection of fresh retained tuples containing only None.

The new python_implementations metadata lets benchmarks declare which Python implementations they support. Selection checks the version and implementation of the target interpreter. Benchmarks that do not support the target implementation are reported and skipped. The frozen and resurrection cases are CPython-only; the large-collection and tuple cases also support PyPy.

Related to python/cpython#159437.