Is there anything besides (massive) engineering challenges standing between more advanced devices like this physically eating away at cancerous tumors in the same way physicians use special maggots for wound debridement?
We just don't have the expertise for those kinds of thing, at least not yet. A plethora of sensors (that can detect cancerous cells) and this sort of technology would be needed. We'd essentially be trying to create artificial cells, a monumental task nature took billions of years in refinement (even if we can find shortcuts sometimes) -- too many very complex systems (energy, better sensors, better movement inside the body, "cognition", etc.).
Maybe in 50 years of continued research? 100 years? I don't know.
[1] The state of the art is about 100 transistors per square micron. This assumes you could drive and interface with them adequately. Cells accomplish all they do without any transistors at all, it's all protein-based sensors and DNA-encoded behavior, I believe.
What you're talking about sounds more like artificial tissues and organs, which is an active area of study, although I don't immediately see how we might "engineer-out" cancer. There exists a world in which we develop tissues that age slower, but the more likely (not that it's happening anytime soon) scenario is very cheap organs that we can swap out for ours.
As for more traditional robots with ~100 transistors, why do they need to be autonomous? Why not mesh network them and control them as drones?
Immunotherapy looks for ways to counteract this 'feature' and allow the immune system to do its job and murder the bastards.
We're still effectively in the early days of this approach to treating cancer and results are mixed, but it shows incredible promise and I feel that it ultimately will be the way we beat most if not all cancers.
Obviously, as you note, there are enormous (or rather, extremely tiny, but very hard!) problems with this. However, it's not all doom and gloom.
I could imagine that going in the 3D dimension for example, could enable another 100X increase in transistor count. In addition, you shouldn't need to actually do the computation onboard for cancer applications, just transmit it via some method, say RF to an outside computer that can do the compute work. Finally, cells are something like 10-30 um in diameter, which gives a lot more room to play with than 1um (the linked paper has scale of about 100um I believe).
"Smart weapons" kill innocent people all the time, no matter how many times we repeat smart in their model name. GPS still lead truck drivers directly to the bottom of a lagoon.
If we can't teach a huge machine full of circuits to distinguish between an ambulance or a tank, how we can realistic trust that tiny machines, with no space at all for complex circuits, will be able to diferenciate between a tumor or our hypothalamus?
What would happen if the patient approach a magnet accidentally? Will the army of nowaybots be displaced enough cm to try to fry your pancreas assuming that they are in a different location?
So they could be injected into a tumor and powered by an external radio source.
In they're only powered by light a fiber optic filament could be injected in the activation site. So long as the light passes through enough tissue to power them you don't need to stick a halogen bulb up someone's butt.
/sarcasm
But that's me speculating, I don't know how to power an injectable micro robot. If I did the link might have been about me! RF might be completely inappropriate for power.
If you could direct a swarm of micro-robots towards the sites of metastasized cancer growths (which kill you because we can't go in and cut them all out without killing the patient from the surgery) then the prognosis would change substantially.
The holy grail would in essence be, having a system which can get to all the places cancer can, at the same sorts of size scales cancer operates - rather then needing to be in the limited range of things we can surgically intervene on.