Shared_writer_fd.close leaves its mutex locked after a successful close

`Buf.Shared_writer_fd.close` acquires `t.lock`, marks the descriptor closed, and calls `Unix.close`. The mutex is currently unlocked only inside the exception handler for `Unix.close`.

On the normal successful path, `close` returns with the mutex still held. Any later `write_fully` call blocks while trying to acquire the lock instead of observing `Closed`. A second `close` blocks for the same reason. This can also leave a writer queued forever if it races with shutdown.

The close error is intentionally ignored, but the lock release should happen regardless of whether `Unix.close` succeeds. I have a small patch ready that puts the best-effort close inside `Fun.protect` and adds a regression covering `close` followed by `write_fully`.

Start at Buf.Shared_writer_fd.close and trace its lock handling around Unix.close and Fun.protect. Verify completion with a regression covering close followed by write_fully on the successful-close path, ensuring later operations do not block.

Written by the indexing model from the issue text.