Add `fsync()` of _directories_ when renaming files to `.committed`

Nobody has claimed this yet.

While reading about atomic rename semantics as part of #7029, we've surfaced something further that we may want to do to ensure the renames succeed, with the contents of the newly created file. Specifically, if we _don't_ `fsync()` the parent directory, it is technically possible for the rename (especially around restarts) to access a version of the directory metadata that doesn't know the file exists. So we should, when dealing with `.committed` files:

1) Before `close`ing a `.committed` file, `fsync()` that file
2) After `close()`, call `fsync()` on the parent directory, to minimise loss window, and be sure it's known before an incoming `rename()` call
3) After `rename()`, call `fsync()` on the destination directory, to minimise the window where the rename could be lost

This probably requires shifting to direct `fopen`/`fwrite` calls, so we have file-descriptor access for `fsync()`, and dropping the current use of `std::filesystem`/`std::ofstream` in many places.

Locate the handling of `.committed` files and the current `std::filesystem`/`std::ofstream` write and rename paths. Trace the file close, parent-directory, and destination-directory lifecycle, then verify that each listed durability point is covered and that existing persistence behavior remains intact.

Written by the indexing model from the issue text.