IRF metrics should not assume linear binning

The IRF metrics that load the IRF tables into a Hist make an assumption of linear binning that may not be correct! There is no good reason to do that, since Hist supports arbitrary binning.

Bad (current code), where e_bins are the bins defined in the FITS IRF table:

axis.Regular(
    bins=len(e_bins) - 1,  
    start=e_bins[0].value,
    stop=e_bins[-1].value,
    name="Energy",
    metadata=dict(unit=e_bins.unit),
    transform=axis.transform.log,
),

Should be replaced with:

axis.Variable(
    edges=e_bins.value,
    name="Energy",
    metadata=dict(unit=e_bins.unit),
),

In fact, a lot of the code in the IRF Metrics could be refactored into a single function that properly loads a standard IRF table into a set of axis bin edges and the value histogram. That function could be fairly generic:

ebins, offbins, value = read_irf_from_table(value_col="EFFAREA")

And that could even automatically return a Hist. I'm not sure I understand the reasoning for turning the offset bins into categories instead of keeping them as non-category axes, so even that could be simplified.