Why does rendering on a PC take 10 seconds compared to 0.25 seconds in-camera?

The camera gets the image directly from the sensor to the buffer memory where it is converted using whatever settings you have chosen and embeds it into the RAW file that is written to the memory card. The whole process does not necessarily complete before the camera is ready to take the next photo - that is why it is called "buffer" memory. Compare this to a PC which has to read the file into memory - if from a slow HDD then you could be looking at nearly 1 second alone to read into memory (e.g. some slow 4200/5400 rpm drives towards the inner tracks or that are heavily fragmented). If you are reading directly off the memory card itself over USB or card reader, then it will be bottlenecked by the slowest link in the chain (be it the card, the reader or the USB interface). Next it has to do the conversion and then has to re-save the resulting image back to the storage medium.

Comparing the actual processing power between the two devices is trickier. One is using a chip custom designed to do a small set of tasks and completely optimised for it. The other is a general purpose chip that has to cater to every whim of any software developer. If it helps to make the point, then consider the difference between a gaming console and a PC - where despite relatively weaker hardware, the console can produce similar results due to being a fixed target where the developers can communicate directly with the hardware vs. going through some API's and other layers of abstraction. Now compare a console like the PS3 to say a dedicated BluRay player and the same scenario plays out again, the PS3 will consume over 70 watts vs under 20watts for a standalone player when watching a movie (the original PS3 consumed nearly 180 watts to do this same task).

Next you have to understand that the conversion that is done in camera using those specialised chips is taking place at a reduced level of precision. You can see this for yourself if you were to set your camera to RAW + JPEG (at the highest quality setting) and process the RAW file in Nikon software without modifying any settings and saving the resultant file to JPEG at a similar compression. The differences are subtle and require pixel peeping in most cases, but it is there nonetheless. For an analogy look to computer graphics cards, where in some cases, the difference in speed going from single floating point precision to double precision can reduce processing speed by a factor of 5. For gaming and other tasks for which precision is not that important and for which speed is, then it is a good trade-off. Since pixel peeping is required to detect differences, then camera manufacturers prefer to put less precise conversion in the camera itself rather than slow down the shooting frame rate. They assume (and rightly so) that anyone who cared about the last few percent of image quality will already be shooting in RAW anyway.

In response to your other points about building a specialised "add-on DSP" or some sort to speed things up, this would only be able to help in some cases - i.e. where the CPU is the primary bottleneck. Given that it is such a specialised bit of hardware, then it would be very costly and people would rather just invest that extra cost in faster "general purpose" hardware. The rationale being that the specialised chip would only offer improvements for a singular task, vs having a faster general purpose processor whose speed would improve everything. As for offloading to the GPU, well that might help, but again, only if the bottleneck were severe enough to offset the whole grab the image from storage medium, push to memory, then push from memory to GPU memory, do the processing, then push back to main system memory to be written to storage. In the computer world, there are many examples of this playing out. e.g. a long time ago, playing back DVDs was very CPU intensive, so many DVD kits came with a dedicated hardware MPEG2 decoder, you don't see these any more as even the slowest PC can play DVD streams.