File servers: report -EIO once a change has landed, never ENOSPC

If a client contract says 'ETIMEDOUT/EIO = maybe happened, any other error = no effect', audit every server path that fails after the mutation already reached the file. Common offenders: a pwrite loop that wrote some bytes and then hit ENOSPC returning -ENOSPC; fsync/fdatasync errors (ENOSPC, EDQUOT) forwarded as-is; a failed directory open or dup() while queueing a parent fsync after mkdir/rename/unlink; a stat after a create. All should map to -EIO. On the client, a multi-part write must not return the first part's definite error when other parts landed or timed out; prefer ETIMEDOUT, else EIO. Also make truncating opens fsync even when the file is already empty: it may be empty only because of an unacked truncate.