Even the cheapest baseband chip will add a significant cost to a phone. If they can get rid of the physical baseband processor they get a major cost savings.
As for why user a hypervisor vs a user mode or kernel driver: Modularity.
Some company can specialize in producing soft-basebands (software eats the world!) and ship a simple image blob that you can drop into your virtualization host, set it to use 5% of CPU time and you're away to the races.
Might be able to bring up a baseband on a new platform inside a day.
(Not that a shanzhai manufacturer making a $34 phone cares in the slightest about those things)
In that case, and with a CPU and peripherals designed with keeping this concept in mind, the overhead will be almost non-existant compared to, for example, running the baseband functionality as a "regular kernel module"; with the added advantage of clearly separating both worlds, and isolating the realtime-requirements of the baseband from the Linux kernel.
For some similar concept, have a loot at "RTlinux", where the Linux-Kernel runs as the lowest-priority task in a very thin realtime-OS to achieve guaranteed reaction times for, e.g. motor control applications.
I really don't understand why this makes sense to do. What is so special and secret about the baseband that makes it worth the effort of running a hypervisor on such an underpowered phone?