escape the hostname every marker publishes

[dxbjavid]

Every marker format in this package holds the hostname to printable non-space ASCII. socket.gethostname() reports whatever the kernel stores: a space, a newline, or a byte no codec encodes, which Python hands back as a surrogate. Publishing one raw makes a holder write a record it cannot read back.

Measured against 3.32.3, on a host named host with space:

• SoftReadWriteLock.acquire_read takes the slot, then the heartbeat stops on its first tick, so the next peer ages the marker out and takes the lock from a live reader.
• SoftReadWriteLock.acquire_write never completes. The writer cannot recognize its own claim, so it waits out the timeout, evicts, rewrites, and loops.
• StrictSoftFileLock.acquire raises SoftFileLockProtocolError, since the claim validator rejects a hostname holding whitespace.

On a host whose name is not valid UTF-8, every writer raises UnicodeEncodeError before the marker exists, the soft read/write state lock included. A newline forges an extra line in a protocol 1 marker, so SoftFileLock.is_lock_held_by_us reports False for a lock this process holds.

Escaping only in the soft read/write writer leaves the strict and protocol 1 formats broken, so the escape lives in host_name(), the one place all four writers read the hostname. ? is illegal in a hostname, so escaping it alongside every out-of-grammar byte as ?<hex> leaves a conventional hostname untouched and still keeps two hosts distinct, which owner_is_stale relies on to refuse to probe a foreign PID.

A marker written raw by an older release on such a host no longer matches the escaped name, so owner_is_stale reads it as foreign and declines to reclaim it, which fails closed rather than detaching a live holder.

Tests pin the escape table, and drive each lock family through its public API on four out-of-grammar hostnames. Reverting host_name() to the raw name fails 11 of them.