btrfs: make find_workspace always succeed

With just one preallocated workspace we can guarantee forward progress
even if there's no memory available for new workspaces. The cost is more
waiting but we also get rid of several error paths.

On average, there will be several idle workspaces, so the waiting
penalty won't be so bad.

In the worst case, all cpus will compete for one workspace until there's
some memory. Attempts to allocate a new one are done each time the
waiters are woken up.

Signed-off-by: David Sterba <dsterba@suse.com>