RDMA/ucma: Do not miss ctx destruction steps in some cases

The destruction flow is very complicated here because the cm_id can be
destroyed from the event handler at any time if the device is
hot-removed. This leaves behind a partial ctx with no cm_id in the
xarray, and will let user space leak memory.

Make everything consistent in this flow in all places:

 - Return the xarray back to XA_ZERO_ENTRY before beginning any
   destruction. The thread that reaches this first is responsible to
   kfree, everyone else does nothing.

 - Test the xarray during the special hot-removal case to block the
   queue_work, this has much simpler locking and doesn't require a
   'destroying'

 - Fix the ref initialization so that it is only positive if cm_id !=
   NULL, then rely on that to guide the destruction process in all cases.

Now the new ucma_destroy_private_ctx() can be called in all places that
want to free the ctx, including all the error unwinds, and none of the
details are missed.

Fixes: a1d33b70

 ("RDMA/ucma: Rework how new connections are passed through event delivery")
Link: https://lore.kernel.org/r/20210105111327.230270-1-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>