These results remind me of this:
Did Pentax implement the dual gain readout? Perhaps so!
The dynamic range in most CMOS DSLR sensors of the 20D era and later is about 13.3-14 stops. This value is from the saturation value (= maximum number of photoelectrons countable before clipping) divided by the read noise (= minimum number of photoelectrons countable before encountering noise)
But why don't we see the entire dynamic range the sensor is capable of? The electronics prevents one from actually extracting it from the sensor. At low ISOs, read noise is high -- meaning the shadows turn noisy, losing the shadow portions of the dynamic range. At high ISOs, clipping lops off the highlight portions of the dynamic range.
I suspect this is why the 7D/5D2 have low ISO noise issues - the read noise is high compared to the maximum signal due to small saturation values (due to small pixels).
The solution posed by the article is to read each pixel twice - one at low gain to not lose highlights, and one at high gain to minimize readout noise. Then, merge the two images. Could this be what Pentax is doing? It certainly seems likely - the sensor's dynamic range (~14.1 EV, according to DXO) is about the same as what would be predicted from using a APS-C sized sensor.
Given the DR from DXO and estimating the saturation value from the D300's sensor, the read noise on the Pentax sensor is roughly 1.8 electrons - a value close to the best that the 1D4 can achieve (1.7e). However, the 1D4 only achieves 11.7EV of DR (according to DXO). This is because of the performance at low ISO, where the 1D4's read noise increases from 1.7e at ISO 12800, to 22.2e at ISO100. The Pentax seems to suffer no such penalty; you have seen this in the RAW file linked in the OP; if you try this with a 1D4 raw, you will get something... ugly.
XTi, 1D3, 2x SB-28, 580EX, 550EX, Tamron 28-75, 50/1.8, "EF" 18-55 II, "EF" 18-55 IS, 85/1.8, 75-300 III USM, 70-200/4L