The memory limitations have been discussed at length at
https://forums.macrumors.com/threads/1698707/ .
Short synopsis is that Intel says that 32 GiB RDIMMs are not supported. Whether they would work is a question.
Alden is correct.
Additionally, there is another important consideration which I do not think has been discussed: the fact that producing a 32GB RDIMM was a lousy idea from a cost/benefit perspective.
The market for large memory models is highly weighted to multi-socket boards (e.g. E5-26xx). These tend to have 3 DIMM sockets per channel. Furthermore, Xeons have limit of 8 memory ranks per channel. Using the most cost effective parts over the past few years one could produce 4-rank 64GB DIMMs.
So, it is technically possible to make a 64GB RDIMM with those DRAM chips but does it make much financial sense? Using RDIMMs would limit these systems to 2 DIMMs per channel (8 total ranks). LRDIMMS use buffering instead or registers. Part of the buffering logic hides internal ranks, making them show up as 2 rank instead of 4 to the cpu's memory controller.
So LRDIMMs make it possible to fully populate these servers since 3 LRDIMMs per channel appear as 6 ranks rather than the 12 ranks that they are internally.
The bottom line, though, is that the bulk of the market for the past few years (HP proliant, IBM, etc.) built systems which had an additional constraint (8 ranks) which made LRDIMMs more viable than RDIMMs. So even if someone made them few would buy them.
Maybe some low-volume, boutique producer currently makes or is testing 64GB RDIMMs. Maybe the existence of the nMP will spur such production. Until then, it is not surprising that people keep seeing LRDIMMs everywhere but no RDIMMs.
p.s. while editing this, I also vaguely remember reading an erratum for the E5 v2 series which said something about a bug with supporting 4 rank DIMMs. I think it involved s workaround for handling a memory sleep state. That might play a part in this as well.