Switch the Pocket EVO to real suspend to RAM

[jesherman]

Switch the Pocket EVO to real suspend to RAM

Depends on the SM8550 deep-suspend patches in armada-packages. Without them
the SoC leaves deep suspend after about seven seconds (TSENS uplow wake IRQs),
so this must not land first.

Two pieces:

ARMADA_SUSPEND_MODE=mem routes the power button to a real suspend instead of
fake-suspend.

SuspendState is pinned to mem. systemd defaults to "mem standby freeze", so a
failed deep suspend falls through to s2idle, which hard-crashes SM8550 -- a
watchdog reset with "PM: suspend entry (s2idle)" as the last line written. A
failed deep should no-op instead.

Measured on a Pocket EVO with a Steam session running, on battery, across four
windows from one hour to five:

1.00h 88,473 uAh = 88.4 mA (capacity 42% -> 41%)
5.22h 481,125 uAh = 92.2 mA (capacity 71% -> 66%)
1.11h 288,169 uAh = 258.7 mA (capacity 62% -> 58%)
6.58h combined = 167.4 mA (capacity 71% -> 58%)

So roughly 90-170 mA, call it 2 to 4 days of standby against about 1.2 for fake
suspend. Deep suspend held every cycle, resumed cleanly, and was woken only by
the RTC alarm, with no suspend failures. The spread between windows is wider
than I would like -- both instruments agree within a few percent on each window
individually -- so the range is quoted rather than a single figure.

That needs the PCIe D3cold series (armada-packages#28/#29). With only the TSENS
and UFS work it is 237 mA: this board carries a discrete Renesas uPD720201 PCIe
USB3 controller for the built-in gamepad whose rails are held by the PCI device
node, so nothing short of PHY power-down drops it.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

---

Depends on armada-packages#28 (validated stack: TSENS + UFS + PCIe D3cold) and armada-packages#29 (its source). Do not merge before them.

[jesherman]

Dropped the Wi-Fi/Bluetooth power-down sleep hook that was in the first revision of this branch.

The idea was sound — NetworkManager only drops the link, so the WCN7850 stays powered with mhi0 wakeup-enabled and both PCIe endpoints pinned at power/control=on, and ROCKNIX's SM8550 sleep.d does a full radio disable — but I could not show it helps. A two-hour run with the hook measured 237 mA; a 45-minute run without it measured 192 mA. Those windows differ in reliability and the configurations differ in other ways, so it is not a clean A/B, but there is certainly no evidence of a benefit, and the hook costs wake-on-LAN and a couple of seconds of resume latency.

Leaving it out until someone can measure a difference over a window long enough to trust.

[jesherman]

@virtudude this is ok to close right? I think anything relevant has since been pulled into sleep testing builds