checklocks: preserve concurrent TCP close updates

Read and write readiness can update the cached closed directions under
different queue locks. Loading the old bits and swapping their union can
lose the other direction's concurrent update. Use an atomic OR instead,
so each update only adds bits, and annotate the atomic-use contract.

All four callers discard the previous state. Remove that unused return
and document that adding the open state leaves the bits unchanged.

Assisted-by: Codex