Improve shortcomings in the first iteration of CKMS

[DieBauer]

After discussion in #755 a next iteration:

• Introduce a dynamic batch size with an upper limit of 500. It looks like smaller batch size gives better results (precision) for single quantiles.
• Skip merging the first item of the samples, to always have the real value of p0.
• Enable the quantile to be 0 or 1, and short-circuit the get method to directly get the values for p0 and p100.
• Fixes a bug that the pXX value was not within error bounds

[fstab]

Fixed with #755.