And definitely read the relevant sections of The Art of Electronics.... If precision is important its vital to design the circuit correctly. You don't want to find you're just measuring the temperature of your electronics!
However if absolute precision isn't important an op-amp stage with adjustable offset might be quite useful in boosting your signal. You mention a resolution of 100uV but don't say what the range of values is, nor if absolute v. relative precision is important.
For serious applications I wouldn't recommend any Arduino board as they all connect analog ground to digital ground - this will lead to a variety of problems in precision measurement.
I have a problem with that line -- the rest I agree...
CrossRoads above pointed out this link.http://www.gravitech.us/i2c128anco.html
Now a 50KHz sampling rate and a choice of single ended or differential output should work for temperature on ovens or monitoring heat loss -- this is an application where the variations are slooowwww -- so you only have to deal with the accuracy.
I am using a Mega2560 to do things which appear to be difficult with some of the other controllers I have here. Each problem has a specific problem space. Your generalization is correct -- but you could convince people to not do something that is done cheaply and effectively on the arduino.
Now I do think that I have proved that 150 samples per second (SPS) is about the best I can do logging data via a UDP collection system. So I am going to caution people about that. But I expect that when I finally post some code people will prove that I don't have a clue what I am doing and that suggestions for improvement will be made. I appreciate that and hope people are that critical. I tried to push it to the 440Hz limit of the BMA180 in Mode 3 at 1200Hz bandwidth -- it just didn't work out. It dropped almost 2 out of three 48Byte UDP packets. A smaller packet might help a bit.
But.... within those bounds I am happy.
Using TWI/I2C and SPI sensors that offload the work make a real improvement in the collection rates possible.
Once you have developed an approach, You have to test to confirme the calculations...
Once you have a base of tests -- then your predictions are likely to be better.
Just my $0.02 which is probably worth less than your $0.02