dmaengine: sdxi: Context management overhaul

This survives booting through dmatest and unbinding the device with
kasan and kmemleak enabled in a VM.

Some key points:

* Define a class for tracking contexts in the device's client_cxts
  xarray, allocating the context IDs, and automatically unwinding on
  early return.

* Ensure valid bit updates for CXT_CTL and CXT_L1_ENT occur after the
  entry is fully programmed, preceded by a barrier.

* Use intermediate logical representations of control structure subfields
  that are passed to the serialization functions.

* More clearly separate treatment of admin and client contexts. Do not
  allocate context ID 0 from sdxi_dev->client_cxts any more.

* Do away with goto-based error unwinding in favor of a robust
  sdxi_cxt destructor that handles partially-allocated contexts.

Signed-off-by: Nathan Lynch <nathan.lynch@amd.com>

Author: nathanlynch