For single-primitive stores (e.g. AtomicBitSet<AtomicU64, V>), each method
operates on one atomic word — all operations are truly atomic with respect to
the entire bitset.
For multi-word stores (e.g. AtomicBitSet<[AtomicU64; 4], V>), each individual
operation (insert, remove, contains, …) is atomic per word, but
the bitset as a whole is not a single atomic unit. Concurrent modifications
to bits in the same word are correctly synchronized. Modifications to bits
in different words are independent atomic operations with no cross-word
transactional guarantee. Composite queries like len(), iter(), or
is_subset() may observe a mix of old and new state across words under
concurrent mutation.