fix error bars in log scale

[Moelf]

fix #2961
before:

after:

[mkborregaard]

I'm trying to understand why you have sub-zero error bars in log space in the first place?

[Moelf]

that's a valid question.

because the statistical error for counting experiment is sqrt(N), if N<=1 (the value is scaled, because of limited statistics in real world), the statistical error is >=N itself. While it's un-physical, currently we can't have asymmetric error bar. And it's often not easy to derive what's "correct" value for the lower half, so this is still a good visual cue to say: don't read too much into this value since its error bar is large.