Fix memory consumption related to SMB AIO reads

• Add backpressure mechanism for AIO reads

In some edge cases where storage backend can't keep up with client
reads, io_uring / samba will allow queueing up reads beyond what
is reasonable leading to OOM errors in some extreme edge cases.

This commit keeps a counter of outstanding read requests per-TCON
and switches to pread(2) once the queue depth limit is reached.
This effectively applies backpressure to client and prevents excessive
memory consumption. A total count of synchronous reads performed is
logged when TCON is disconnected if log level for the VFS module is
sufficiently high (3 or higher).

• Fix memory leak related to AIO memory pool for some failure scenarios
  for AIO reads.

The memory pool for AIO buffers is allocated under a memory context
that persists for the duration of the SMB session. When chunks of
pool are assigned out for a particular AIO request, a separate
small allocation (io_link) is performed with a pointer to the data
buffer and a talloc destructor set such that when the io_link is
freed the buffer is returned to the pool.

Initially these were performed as two separate steps (assigning out
buffer from pool and setting up io_link) which allowed for situations
to arise in which the request could error out before the io_link was
set resulting in the buffer never being returned to the memory pool.

This commit revises the workflow and API for the memory pool such
that the chunk is always assigned with an associated io_link allocated
under the memory context of the initial SMB2 read request. Once the
request is successful and we're getting ready to respond to client
then we reparent the io_link to the context of the response. This
ensures that the buffer is always returned when a limited-life
talloc chunk is freed.