Basically it's just up to everyone to test stuff however they want to. The scheduler people test schedulers, the intel wifi driver people test intel wireless drivers, IBM tests has someone running the Linux Test Project.
Since it's all decentralized there aren't any test farms. It's just random people testing on random computers that they have. For example, I have a computer that I use for fuzz testing the latest kernels. Other people have similar setups.
But really the most important thing is to get regular users to run the latest kernels and report bugs. We try to make it as easy as possible to install the new kernel although it is a bit complicated to configure it.
If a kernel hasn't been released yet, and you have a bug and you know which cset introduced it then you can report that to LKML. The author of the cset has a few days to figure out a fix, otherwise it gets reverted. But mostly people just know there is a bug and don't know which cset introduced it. Also once a kernel has been released, it is sometimes harder to do the revert. Maybe there is a program which relies on the new behavior.