The prepare_transfer_hardware() is called in atomic context and
calling synchronous runtime pm calls can create scheduling deadlock.

Therefore, in place of calling runtime PM calls from prepare/unprepare
message transfer, calling this in transfer_one_message().
Signed-off-by: Laxman Dewangan <ldewangan@nvidia.com>
Tested-by: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>