allocator: add CCO-backed LSA symmetric windows

[yangyuhuiling]

Summary

• preserve PyTorch's pointer-array contract while deriving peer pointers from a CCO LSA window
• add LSA-only communicator setup, rank-local window identity, topology introspection, and GC-safe resource lifetimes
• expose an owning Python LSA handle with divergent-slot, subgroup, teardown, and FlyDSL example coverage

Test plan

• BUILD_TORCH_SYMM=ON python setup.py build_ext --inplace
• pytest -v tests/python/allocator (4 passed)
• CCO host test (8 ranks) and divergent-slot regression (2 ranks)
• torch-to-CCO import and FlyDSL LSA all-to-all examples (2 ranks)
• pre-commit checks on all changed files

Made with Cursor