gh-158525: Avoid reacquiring dict lock when copying

[n1snt]

When copying a sparse mutable dict, PyDict_Copy() already holds the source lock. The compaction fallback called dict_merge(), which tried to acquire the source again as part of a two-object critical section. In a free-threaded build, this caused the thread to spin and park on its own lock.

Call dict_dict_merge() directly when the source uses the built-in iterator. The destination is newly created and not visible to other threads. Keep the generic merge path for frozendicts and subclasses with custom iteration.

The existing test_copy_noncompact test covers the changed copy path.

Benchmark

Linux arm64, free-threaded current main release build:

Operation	Before	After
compact.copy()	0.108 us	0.108 us
sparse.copy()	9.86 us	0.169 us
dict(sparse)	0.178 us	0.173 us

Tests

• Free-threaded debug build: ./python -m test -v test_dict test_free_threading.test_dict
• GIL-enabled debug build: ./python -m test -v test_dict

AI use: I used an AI assistant to help inspect the locking path and prepare the change. I reviewed the diff and ran the tests and benchmarks above.

Fixes #158525

• Issue: Free-threaded dict.copy() spins on an already-held source lock when compacting a sparse dict #158525

[python-cla-bot]

All commit authors signed the Contributor License Agreement.