There are a few things that I find difficult to understand about this.
Firstly, a freely oscillating circuit inside the CPU to generate random bits doesn't seem complicated to me. It should extract true randomness as fast as you could move bits from the CPU to memory. Why hasn't it been done?
Lacking a CPU instruction, virtual systems could just have a random driver to help them along. They would receive data from the host, using whatever combination of true random / pseudorandom generation as necessary. It's in all likelihood of low complexity problem, an afternoon's hack. Why hasn't this been done?
It only takes in the hundreds of bits range of true randomness to properly seed a random number generator, after which it can be run for significant periods (probably years?) without leaking the prng state. In fact, if you unpredictably flip the bits in the prng state based on true random inputs, the problem seems entirely solved to me, as no amount of reading the prng output reveals anything consistent about its state. Am I missing something, again? (Hypothetical answer: people simply are really paranoid about randomness of their random numbers. But is such paranoia justified?)