We currently insist on disabling PAuth on vcpu_load(), and get to
enable it on first guest use of an instruction or a key (ignoring
the NV case for now).

It isn't clear at all what this is trying to achieve: guests tend
to use PAuth when available, and nothing forces you to expose it
to the guest if you don't want to. This also isn't totally free:
we take a full GPR save/restore between host and guest, only to
write ten 64bit registers. The "value proposition" escapes me.

So let's forget this stuff and enable PAuth eagerly if exposed to
the guest. This results in much simpler code. Performance wise,
that's not bad either (tested on M2 Pro running a fully automated
Debian installer as the workload):

- On a non-NV guest, I can see reduction of 0.24% in the number
  of cycles (measured with perf over 10 consecutive runs)

- On a NV guest (L2), I see a 2% reduction in wall-clock time
  (measured with 'time', as M2 doesn't have a PMUv3 and NV
  doesn't support it either)

So overall, a much reduced complexity and a (small) performance
improvement.
Reviewed-by: Oliver Upton <oliver.upton@linux.dev>
Link: https://lore.kernel.org/r/20240419102935.1935571-16-maz@kernel.orgSigned-off-by: Marc Zyngier <maz@kernel.org>