They have certainly reverse-engineered some variants of Intel 486.
Nevertheless, they have implemented them in different and better manufacturing processes, so these are not copies of the Intel CPUs, any more than the Compaq computers would have been copies of the IBM computers.
They could have done the reverse-engineering from just buying some CPU samples from any computer shop. It was hard work, but still many orders of magnitude easier than reverse engineering a CPU today.
They could have also stolen some 486 mask sets, which would have made the reverse engineering easier, but this was not necessary, it would have been just a cost-saving measure, with high risks and modest benefits.
Wouldn't it be possible to use the same masks to get a basic "shrink", at least for these rather primitive late 80s / early 90s processes? Likely impossible today because of everything being heavily tuned to the wavelength of the lightsource.
It depends on how different were the manufacturing processes.
Newer improved processes usually have some extra masks that would need to be generated, besides shrinking the masks of the original set.
Also, the original masks may need to be shrinked non-uniformly, i.e. by modifying the size ratio between holes and covered areas, not by a simple geometric shrink.
Even when the new masks could be generated mostly by shrinking, there is no chance to make such a complex device work in the new manufacturing process without reverse-engineering it. Correct behavior may depend on timing relationships that would be changed in the new CMOS process and which could require small schematic changes for compensation.
After the schematics is extracted by reverse-engineering from etched CPU samples or from a set of masks, it is probably as easy to draw again the masks as to shrink the old masks, even if the old masks or CPU samples would certainly be used as guidance for the floor plan.
Chinese company, so sure industrial espionage would be the first guess. But to be fair lots of folks reverse engineered the 486 (as shown in TFA). If the performance analysis it to be believed, something was improved, so it's not an exact copy.
