[pulls] add task to compare nuisane pulls and impacts from multiple datacards (e.g. categories)

Hi @mrieger,

This PR adds a new task, which allows to compare Pulls and Impacts for multi-datacards:

law run PlotMultiplePullsAndImpacts --version test --multi-datacards top1.txt:top2.txt --datacard-names "Top resolved,Top boosted" --keep-failures --campaign 2016 --PullsAndImpacts-workflow htcondor --workers 2

yields: multipulls.pdf

This task comes in very handy, when debugging nuisance parameters, which e.g. are pulled in opposite directions for different categories, and thus create tension in the fit.

I've tried to keep the code changes minimal, by reusing the already well-defined plot function.

Unfortunately I was not able to add docs, since the example datacards seem to be broken for me right now. However, I can add a task description and an example command to the docs if you like.

Best, Peter

dhi/plots/pulls_impacts.py

@@ -125,7 +125,7 @@ def plot_pulls_impacts(
         print("{} remaining parameters after filtering".format(len(params)))
 
     # apply ordering
-    params.sort(key=lambda param: param.name)
+    params.sort(key=lambda param: (param.name, param.tag))
     if order_by_impact:
         # failures to np.inf
         v = lambda x: np.inf if np.isnan(x) else x
@@ -245,7 +245,8 @@ def plot_pulls_impacts(
         h_dummy.GetYaxis().SetBinLabel(1, "")
         for i, param in enumerate(_params):
             # parameter labels
-            label = to_root_latex(labels.get(param.name, param.name))
+            label = param.tag or param.name
+            label = to_root_latex(labels.get(label, label))
             if param.invalid:
                 label = "#bf{{{}}}".format(label)
             label = ROOT.TLatex(x_min - (x_max - x_min) * 0.01, n - i - 0.5, label)
@@ -365,6 +366,10 @@ def plot_pulls_impacts(
         # campaign label
         if campaign:
             campaign_label = to_root_latex(campaign_labels.get(campaign, campaign))
+            # if we have a 'tag', we only plot the tag name, so we add the actual nuisance name once
+            # together with the campaign
+            if param.tag is not None:
+                campaign_label += ", {}".format(to_root_latex(labels.get(param.name, param.name)))
             campaign_label = r.routines.create_top_left_label(campaign_label, pad=pad, x_offset=22,
                 y_offset=90 if paper else 105)
             draw_objs.append(campaign_label)
@@ -410,6 +415,7 @@ class Parameter(object):
         self.type = data["type"]
         self.groups = data["groups"]
         self.invalid = np.any(np.isnan(self.poi))
+        self.tag = data.get("tag")
 
         # compute two sided impacts, store as [low, high] preserving signs
         self.impact = [