ctools

Formally I agree, I added this feature when working on #2312.

Originally the code was:

# Perform maximum likelihood fit
like          = ctools.ctlike(obs)
like['edisp'] = self['edisp'].boolean()
like.run()

# Continue only if log-likelihood is non-zero
if like.obs().logL() != 0.0:

but I then switched to

# Perform maximum likelihood fit. Skip bins that have no events.
like          = ctools.ctlike(obs)
like['edisp'] = self['edisp'].boolean()
if like.obs().nobserved() > 0:
    like.run()

# Continue only if log-likelihood is non-zero
if like.obs().logL() != 0.0 and like.obs().nobserved() > 0:

Unfortunately I have no track why exactly I did the change, I think it was to avoid fitting errors. I have no problem with putting this back again, but we should do some testing to check that there are no side effects.

I put back the original code.

Code merged into devel.