Unlock shared writer after close

[jeet-dekivadia]

Fixes #25.

Shared_writer_fd.close acquired the writer mutex but released it only when Unix.close raised. On the usual successful path, later writers waited forever before they could observe Closed.

Wrap the best-effort close in Fun.protect so the mutex is always released. The regression closes a shared writer and then checks that a later write raises Closed instead of blocking.

Validation:

• git diff --check
• private Linux runner: installed OCaml 5.2.1 and ppx_jane, then ran dune build @all -j2; compilation reaches the repository sources but the public toolchain cannot compile existing Domain.Safe and Sys.Safe uses and existing warning-as-error findings, so dune runtest is not reachable outside the Jane Street build environment

[jeet-dekivadia]

Small follow-up here: this is scoped to the close path after the writer mutex is acquired. On the successful Unix.close path the mutex was not released, so later writers could block before observing Closed.

The patch wraps the close attempt in Fun.protect, preserving the existing best-effort close behavior while making unlock unconditional. The regression checks that a later write raises Closed instead of hanging. DCO is green.

[jeet-dekivadia]

I rechecked this today while it is waiting for review. The ownership point is the writer mutex after Shared_writer.close has acquired it: successful Unix.close used to leave the mutex held, so later writes could wait instead of observing Closed.

The patch only makes unlock unconditional around the existing close attempt; it does not change the closed-state transition or the best-effort close behavior. DCO is still green.