RDMA/srp: Make struct scsi_cmnd and struct srp_request adjacent

Define .init_cmd_priv and .exit_cmd_priv callback functions in struct
scsi_host_template. Set .cmd_size such that the SCSI core allocates
per-command private data. Use scsi_cmd_priv() to access that private
data. Remove the req_ring pointer from struct srp_rdma_ch since it is no
longer necessary. Convert srp_alloc_req_data() and srp_free_req_data()
into functions that initialize one instance of the SRP-private command
data. This is a micro-optimization since this patch removes several
pointer dereferences from the hot path.

Note: due to commit e73a5e8e ("scsi: core: Only return started
requests from scsi_host_find_tag()"), it is no longer necessary to protect
the completion path against duplicate responses.

Link: https://lore.kernel.org/r/20210524041211.9480-6-bvanassche@acm.org


Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>