pmdomain/rockchip: skip QoS operations for idle-only domains

The NanoPi Zero2 (RK3528) kernel panics during boot when a GPIO-controlled USB VBUS regulator is defined on GPIO4 (which is in PD_RKVENC). The goal of this series is to make USB host power work on boards that use GPIO4 for regulator control.

The root cause is a probe ordering issue. On RK3528, the power domain controller's first probe attempt fails because PD_GPU's clock lookup returns -EPROBE_DEFER (CRU hasn't probed yet). The driver then tears down all domains, including PD_RKVENC which would have registered successfully (it has no clock requirements). During this window, the USB regulator driver probes and requests GPIO4, which is in the now-unregistered PD_RKVENC -- this triggers a synchronous external abort.

With patch 2 alone (skipping deferred domains), the idle-only domains register successfully. But the genpd framework then attempts to power them off via genpd_power_off_work_fn. This calls rockchip_pd_power(), which does QoS save and idle requests on domains with pwr_mask == 0 that cannot actually be powered off.

To your question about why QoS registers become inaccessible on idle-only domains: I have not root-caused that specifically. What I can confirm is the crash trace below, which occurs when patch 2 is applied without patch 1. The abort happens during rockchip_pmu_set_idle_request on an idle-only domain:

  Internal error: synchronous external abort: 0000000096000010
  CPU: 2 PID: 60 Comm: kworker/2:3
  Workqueue: pm genpd_power_off_work_fn
  pc : regmap_mmio_read32le+0x8/0x20
  lr : regmap_mmio_read+0x44/0x70
  Call trace:
   regmap_mmio_read32le+0x8/0x20
   _regmap_bus_reg_read+0x6c/0xac
   _regmap_read+0x60/0xd8
   regmap_read+0x4c/0x7c
   rockchip_pmu_set_idle_request.isra.0+0x94/0x1b4
   rockchip_pd_power+0x37c/0x608
   rockchip_pd_power_off+0x14/0x38
   genpd_power_off.isra.0+0x1f0/0x2f0
   genpd_power_off_work_fn+0x34/0x54

The two patches work together: patch 1 prevents QoS access on idle-only domains, and patch 2 prevents the full probe teardown when a single domain defers.

Tested on NanoPi Zero2 (fixes panic) and Radxa E20C (no regression).