md: delay choosing sync action to md_start_sync()

Before this patch, for read-write array:

1) md_check_recover() found that something need to be done, and it'll
   try to grab 'reconfig_mutex'. The case that md_check_recover() need
   to do something:
   - array is not suspend;
   - super_block need to be updated;
   - 'MD_RECOVERY_NEEDED' or 'MD_RECOVERY_DONE' is set;
   - unusual case related to safemode;

2) if 'MD_RECOVERY_RUNNING' is not set, and 'MD_RECOVERY_NEEDED' is set,
   md_check_recover() will try to choose a sync action, and then queue a
   work md_start_sync().

3) md_start_sync() register sync_thread;

After this patch,

1) is the same;
2) if 'MD_RECOVERY_RUNNING' is not set, and 'MD_RECOVERY_NEEDED' is set,
   queue a work md_start_sync() directly;
3) md_start_sync() will try to choose a sync action, and then register
   sync_thread();

Because 'MD_RECOVERY_RUNNING' is cleared when sync_thread is done, 2)
and 3) and md_do_sync() is always ran in serial and they can never
concurrent, this change should not introduce any behavior change for now.

Also fix a problem that md_start_sync() can clear 'MD_RECOVERY_RUNNING'
without protection in error path, which might affect the logical in
md_check_recovery().

The advantage to change this is that array reconfiguration is
independent from daemon now, and it'll be much easier to synchronize it
with io, consider that io may rely on daemon thread to be done.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20230825031622.1530464-4-yukuai1@huaweicloud.com