The entropy generated on those Intel processors is based on the effects of thermal noise on a simple digital uncertain circuit. That uncertain circuit is part of the processor. Previously processors used an external analog mechanism but newer generation CPUs get these zeroes and ones from that internal circuit. Effectively you have a generally deterministic CPU with a non deterministic component.
Regarding PRNGs many of them have been studied and put through tests of randomness (ie. Die Hard test, NIST randomness tests). Some algorithms are much better than others. The better ones for the purposes of Backgammon are virtually indistinguishable from a real source assuming they are coed properly. The primary problem with many PRNG algorithms is the issue of predictability which make many of them unsuitable for highly secure crypto applications.
Some servers over the years have used external sources for randomness. Usually ones that hooked up to serial and USB ports.
As for this statement In some cases, they do some manipulation to "normalize" results, like limiting the number of 66 in a row. this is possibly true. We had a case of this on Safe Harbor Games a number of years back. Neil Robbins on this forum had posted some basic stats that suggested that the random dice on that server were not so random. I was originally sceptical and wrote some simple scripts to analyse the dice from SHG on matches I already played. 10 times the sample Neil had. Shockingly at the time it confirmed Neil's data. William Womack and I then spent weeks collecting a couple million rolls while watching matches. A pattern emerged. Effectively the occurrence of all doubles was lower than expected. When the developer of the site was asked about it they confirmed they used Mersenne Twister but artificially lowered the doubles thinking it wouldn't be detectable. It was. That is why that site has red dice and yellow dice rooms. The red dice are the older manipulated dice and the yellow manipulated ones.
To this day there are people who think the red dice (the manipulated ones) are fairer and better representative of real world dice. To them the lower number of runs of doubles is perceived as more random in their mind because the games seem to be fairer.
I have a rule of thumb now. If someone says the dice on a server seem to be fairer than other sites then my mind tosses up a red flag. My mind will now assume that such a site probably has dice that do not represent perfectly weighted 6 sided dice. The less complaining, the more likely there is a problem in my mind.
