A very common usage for /dev/mem is to access a device through memory mapped IO from user space.

Cacheable accesses would break that use case.

If my memory serves me correctly you can control the mapped caching attributes using file flags, like O_SYNC, etc. Might also be architecture dependent. If you want to manipulate a memory-mapped device then you don't want caching, but if you want to see memory the same way a typical application does, then you want to keep the caching attributes in your mapping. It's a low-level tool, so there is no right or wrong really.

...it's memory. Where would you cache memory access, in other memory ?

CPU got cache for caching memory