First of all, this is done for the visual effect, not for performance reasons.
The standard way of dealing with performance in physics heavy games is to scale the number of objects, scale the object complexity, and fiddle with engine settings to scale between simulation precision and performance. If there is problems you'd drop what you perceive to be the least significant features.
Do remember though, the industry has made pretty realistic car games for the past ~15 years, with modern computers it's not like they have to scale back to 3 wheels to get things running.
It is true that a collision may cause extra work, how much depends a lot on the specifics of the game, a more detailed physics engine will have a lot of small collisions between different parts that may constitute a significant increase in required computation. But that should be taken into account when the physics is scaled, it's not a problem to get good physics that can still handle some collisions.
If you simply run the physics simulation slower to get slow motion the load will drop proportionately. However, one should note that the requirements for slow motion and real time physics are different, you can afford to have lower precision when stuff happens at racing speed. As long as the player does not notice that the physics engine is wrong it is not a big problem, the slow motion makes the slips much easier to catch, thus the slow motion has a higher precision requirement.
One may choose to use the same physics, scaled to meet both sets of requirements. This solution will require some extra processing power, but it's easy to implement and given modern computers perfectly viable.
Switching physics settings is more complicated, but can potentially result in some gorgeous collisions, not only can one increase the precision, it is also possible to switch the physics models of the cars for better detailed ones that break in more realistic fashion. This mode should end up using approximately the same amount of CPU time for physics as the normal mode, simply because they are both scaled to run at the same minspec configuration.
A middle way is to use a variable step physics engine, those will in general increase the precision when you slow down the simulation, thus solving at least part of the problem. There are other reasons not to use variable step physics, but variable step is still pretty common in the industry.