For me the real question is how good your estimates are. If they are wild-ass guesses nothing will really save them.
Anyway, I don't like:
Just the difference between the Most Likely and the average of the two. It seems like some magic formula which I can't really support with reasonable data.
PERT. It should probably increase your original estimates a bit but this approach still bases just on some estimates and not any hard data.
Other two seem better to me but they also are more tricky:
Standard deviation. To learn what your standard deviation is you need to use historical data for tasks you've completed. For this sole reason it's a better method than two above.
Monte Carlo. With this one I'm not sure what your approach is. Again this is a method which bases on some hard data, so I expect you make Monte Carlo simulation using you historical estimates and real working times. If so why don't you just base estimates themselves on Monte Carlo simulation instead of just calculating buffer this way?
If I'm not missing something you do need to use historical data (estimates versus real working times) at least in a couple of proposed methods. If so, my idea would be to use Monte Carlo simulation against your estimates to calculate new, better estimates which include your track record. There's a great article about evidence based scheduling which describes how you may do it.
In this case I would work on schedule you feel good with (whatever probability level makes you feel good). Then if you need some additional safety catch I'd add buffers which are based on schedule with higher probability of success. For example, after Monte Carlo simulation you get following results: 70% probability that you complete work package in 390 hours and 90% probability it'll happen in 475 hours so you have 85 as your buffer.
If you want to split this buffer into a couple of parts after some work is done simply split your work package into smaller parts and do the same analysis for both parts independently and add the result at the end of each one.
Note: I would prefer just to go with 90%-probability schedule instead of adding buffers to less probable one.
I use approach pretty similar to evidence based scheduling and it proved to deliver pretty good results as long as historical data is reliable so that would be my method of choice.