Add forced BeeGFS client disconnect to fail I/O and unblock D-state processes

[iamjoemccormick]

Originated from #81.

To allow a clean unmount of the file system on client nodes where processes are blocked in BeeGFS I/O due to server-side services that cannot be recovered quickly, it is proposed to add a forced client disconnect mechanism.

When triggered (for example via echo 1 > /proc/fs/beegfs/*/force_disconnect), the client would immediately fail all outstanding and new BeeGFS operations with an I/O error, allowing blocked processes to exit D-state and the file system to be unmounted cleanly.

One argument against this is there are timeouts in BeeGFS already which should eventually unblock processes. However in some scenarios this can take 10 minutes (by default), which is a long time to wait.

[iamjoemccormick]

A better solution might be to implement support for a forced unmount. Some initial notes from the developer who looked into this:

We would have to implement file_system_type::umount_begin() , but we only implement normal kill_sb() for the standard unmount path. umount_begin() is called by VFS when it receives MNT_FORCE flag.

https://git.ti.com/cgit/keystone-linux/linux/commit/fs/namespace.c?id=8b512d9a88875affe584bb3d2a7a235f84343b9e
https://github.com/torvalds/linux/blob/master/fs/nfs/nfs4super.c
https://github.com/torvalds/linux/blob/master/fs/nfs/super.c#L754

Maybe it's the better way to integrate with MNT_FORCE, which will allow synchronous (safe) disconnect + unmount. A /proc knob could be made to be a state transition to "permanent disconnected" as well but who knows, there could be more complications arising from that

This is also related to #72, but that issue is for allowing an unmount to happen when there are no open BeeGFS file descriptors and the management is unavailable, and this issue adds the requirement to allow an unmount to be forced even if there are blocked processes trying to access any service (which would also block a regular unmount).