A cryptographic PRNG would indeed be better than hardware generators, but to be fair we haven't actually proved that the PRNGs we're using are actually cryptographically secure. We're just really sure, though probably more sure than we are that the hardware RNGs haven't screwed up somewhere in the implementation.
Since we need to seed cryptographic PRNGs with something we should try to use something reasonably random and in that sense a hardware RNG is helpful. But I'll agree you don't need megabits of random data to produce secure output and sampling a few timing interrupts will do. Or just use /dev/urandom.