1. The size of the data is often much smaller - the savings from using an nlogn time algorithm vs a n^2 algorithm are much smaller when n=10 versus n=10000. There's a good chance that the constant time setup will dominate for those smaller datasets.
2. To a certain extent, embedded systems don't need to be as fast as possible - they just need to be "good enough" to meet their deadlines so that the rest of the system functions as expected. It'd be nice to be able to compute 1000 CRCs per millisecond instead of only 10, but if it already takes 1 millisecond to transmit the CRC over the serial bus, it's not going to make the system substantially better at its job.
So yes, there is a lot of handcrafted code, but quite often it's just implementing your standard naive brute-force algorithm with minimal code size, minimal memory overhead (preferably without dynamic allocation), and well-understood worst-case latency.