Avoid redundant garbage collector traversals #159437

This issue tracks a series of GC optimizations that we have made in OpenAI’s CPython fork and would like to bring upstream. They target the garbage collector in GIL-enabled builds and reduce repeated walks over the same objects during collection.

The work includes:

• Combining reference initialization and subtraction for full collections, with the permanent-generation bookkeeping needed to preserve gc.freeze() and gc.unfreeze() behavior.
• Skipping finalizer scans and late weakref cleanup when earlier collection phases establish they are unnecessary.
• Clearing unreachable flags and counting unreachable objects during existing passes.
• Skipping reachability traversal for surviving exact tuples that can be untracked immediately.

In production at OpenAI, we've observed roughly a one-third reduction in GC time with these combined improvements.

The changes will be proposed in focused PRs against main, with correctness and performance measurements for each.