There are two issues that I see here:
1) Audio input via USB.
Audio input can be done using android 3.2+ and libusb but it is not easy (You will need to get the USB descriptors from libusb, parse them yourself and send the right control transfers to the device etc). You can get input latency via USB in the order of 5-10 mS with some phones.
2) Audio out in real-time.
This is a perennial problem in Android and you are pretty much limited to the Galaxy Nexus at the moment if you want to approach real-time (using Native Audio output). However, if you master the USB you may be able to output with less latency as well.
I suppose if you go to the trouble of getting the USB to work, you can get a USB audio device with stereo in. If you had connected one mono mic to each of the input channels, then output via USB you would be very close to your stated goal. You might like to try "USB Audio Tester" or "usbEffects" apps to see what is currently possible.
In terms of coding the mixing and output etc, you will probably want one thread reading each separate input source and writing to a queue in small chunks (100-1000 samples at a time). Then have a separate thread reading off the queue(s) and mixing, placing the output onto another queue and finally a thread (possibly in native code if not doing output via USB) to read the mixed queue and do output.
The following Link http://code.google.com/p/loopmixer/ has a flavor for dealing with the audio itself.