dwc2: ship the SOF hold-off (0062) off by default

[mcfbytes]

Makes the dwc2 SOF hold-off (0062) ship off by default, based on a non-RT 6.18.54 measurement. Refs #205.

Evidence

Measured on the 6.18.54 image with the series, toggling the hold-off live through debugfs. Setup: hub, wired pad, BT dongle and an rtw88_8822bu USB WiFi adapter, with iperf3 over WiFi. CPU0 cost is the median time of a fixed CPU0 workload, because this kernel has no IRQ time accounting.

Load	CPU0 workload, on vs off	complete-split misses, on vs off
idle	1,380 vs 1,384 ms	0 vs 0
WiFi send, 126 Mbit/s	4.7–5.0 vs 5.2 s	same
WiFi receive, ~195 Mbit/s	8.6–13.8 vs 5.1–9.5 s (worse in 6/6 pairs)	673–730 vs 553–660 per 40 s

Off RT the hold-off gains nothing when idle and regresses under heavy USB receive. RT was never tested under bulk load. So it starts off, and the debugfs switch stays for testing: echo 1 > /sys/kernel/debug/usb/ffb40000.usb/sof_holdoff.

Change

• 0062 no longer sets hsotg->sof_holdoff at init (the field is zero-initialised, so it starts off). The patch was regenerated from a git tree and its message records the numbers.
• With the hold-off off, the primary handler always wakes the IRQ thread. That's the "M1 off" baseline already measured on RT (13.2%).
• docs/dwc2-usb-irq.md: the 6.18 table, test notes (ARP flux with Ethernet and WiFi on one subnet), two new ledger rows (the bulk-receive investigation, and a 6.18 A/B for 0058), and the upstreaming section now offers nothing yet.

Verification

• The de10nano series replays at -F0 on pristine 6.18.54 (52/52), and the beta series on 7.2.8 (54/54).
• drivers/usb/dwc2 builds W=1-clean from both.
• lint-kernel-patches.sh passes.
• Not booted: the "off" behaviour is the code path already measured on both kernels.

🤖 Generated with Claude Code