📝 docs(util): correct write and break guarantees

[pralav-25]

We documented write_all as keeping the marker "atomic in the process and kernel view". A peer that reads between two os.write calls sees a prefix of the record, and SoftFileLock waits two seconds before it breaks a malformed marker for that reason. I reworded the docstring to say the loop completes the record and each caller has to tolerate partial reads.

We wrote in the break_lock_file docstring, and in the comment in test_break_lock_file_preserves_file_when_mtime_advanced, that keeping a live marker under its break name prevents two holders. Processes A and B both judge a dead holder's marker stale, and B breaks it and acquires. A renames B's live marker aside and, seeing a new inode, leaves it under the break name; that empties the lock path for C to acquire while B runs. B and C both hold the lock. StrictSoftFileLock has no such race because an operator's force_break is its sole way to remove another process's claim, as the strict claims section of docs/concepts.rst describes.

We referred to "the soft read/write marker break" three times in the break_lock_file docstring, but _soft_rw/ has no rename-based break, so I removed those references.

I changed docstrings and one test comment, so I labeled the PR skip news in place of adding a changelog fragment.