Remove `O_RSYNC`, `O_SYNC`, and/or `O_DSYNC`?

[sunfishcode]

These flags are all about providing data integrity guarantees, however to my knowledge, no one has yet investigated the degree to which these guarantees could be made in a portable manner. This suggests that perhaps we should remove them, until we have an idea of what we're actually able to guarantee.

Here are some notes from my survey of documentation I could find:

Very few OS's claim to support O_RSYNC. Some claim to support O_SYNC and many claim to support O_DSYNC or something that sounds equivalent to it, like FILE_FLAG_WRITE_THROUGH on Windows. So if we're going to support them portably, it's important to have a way to implement them without OS support if needed. That raises some questions:

• Is doing an fsync after every write sufficient to implement O_SYNC if a host OS doesn't support O_SYNC? In theory it's not identical because there is a window between when a write has returned and other processes are observing the newly written data, but it's not written to persistent storage yet, so an abrupt power failure may undo changes that had temporarily looked like they had been written. Does that mean we can't implement O_SYNC this way?

• Is doing an fdatasync after every data write sufficient to implement O_DSYNC? In theory it has the same problem as O_SYNC.

• If doing an fsync (for O_SYNC|O_RSYNC) or fdatasync (for O_DSYNC|O_RSYNC) before every read or data read sufficient to implement O_RSYNC? Does an fsync or fdatasync on one file descriptor flush data written through an independent separately-opened file descriptor?