The problem is that with short haul serial links there is a very small operating range where error correction works. Basically the difference between a working link (very low BER) and a link where even ECC is broken (high BER) is quiet small and adding ECC overhead is deemed an overall loss. These links have quite different characteristics from long-haul links that are dominated by ISI. One way to see this is to look at a waterfall plot and note how steep it is on short interconnects. There is a very small range between working fine and completely broken.
Aren't checksums for error detection and not error correction? As I understand it, error correction is having enough information to correct the error while error detection is only knowing an error has occurred.
Error correction only makes sense if the latency of the read/write is long(i.e. spinning physical medium, co-sharrd radio channels).
FEC isn't free(~2x overhead per bit recovery if I remember right) so if your error rates are infrequent then it's worse to use FEC over just resend/reread.
I guess checksums are ok for many applications if USB implements resending of data at the protocol level. For high-speed cameras it might be problematic though.