scadm shouldn't hardcode legacy virtio 9p device ID for SoftNPU

[sion42x]

scadm finds propolis SoftNPU 9p devices by matching device ID 0x1009 with a legacy subsystem ID. Changing subsystem ID to single out SoftNPU from other 9p devices breaks scadm's lookup and vio9p binding.

Moving the device to transitional or modern virtio with device ID 0x1049 will also break lookup. Both of these things came up as blockers in oxidecomputer/omicron#11237.

scadm should discover a device without depending on PCI identity. Similar to how the aforementioned PR approaches it, we can walk devinfo for virtio 9p nodes that accept both 0x1009 and 0x1049, and identify SoftNPU by its version string (9P2000.P4) with a Tversion exchange.

There may be better ways of doing it as well. As @iximeow mentioned to me in passing, there's nothing inherently wrong with it being a Legacy ID. But that plus hardcoding and intolerance of subsystem ID changes makes it a blocker.

[sion42x]

I realized this perhaps should have gone to oxidecomputer/softnpu, but the problem really manifests here. If there's a preference let me know.

[iximeow]

well... the issue about scadm ought to be on softnpu, yes :) and that in turn is a blocker for changing 9p away from the legacy interface because if we went and "fixed" this in Propolis today we'd just be breaking dependencies for no reason

to be clear for future readers: we'd like to be able to move 9p to at least transitional (for consistency if nothing else), but because of the issue above we'll need to make sure scadm has been adjusted to tolerate the future when we do.