bbtautau v6: control channels, background-only fit and faster evaluation

[SimonSPD]

This pull request adds the functions that the bbtautau v6 analysis uses. It also corrects the ML export and makes the evaluation of expressions faster.

New functions

• FriendConfig.channels: ShapeSmith reads a friend only in the listed channels. The default is all channels. A channel without this friend production does not look for the friend files.
• Analysis.control_channels: datacards and fit add the control channels to each final state. A control channel is never a final state of its own. Example: em as the ttbar control channel of the tau channels.
• target of DataMinus and ABCD: the estimator writes its output to this region. The default is the nominal region. Example: QCD in a diagnostic region.
• shapesmith fit --background-only: this command fits the data with r set to 0 and frozen (combine -M FitDiagnostics). It writes datacards/background_fit.{json,md}. The summary shows the fit status, the pull and constraint of each
  nuisance, and the prefit and postfit yields.

Corrections

• The ML export writes for each channel only the processes and estimator outputs of that channel. Before, it could write data for a process that the channel does not have.
• ml-export does not export the control channels.
• The background-only fit uses --cminDefaultMinimizerStrategy 0 --robustHesse 1. Before, Minuit forced the covariance to be positive definite when normalisations were degenerate. The fit then showed uncertainties that were too small.

Performance

• expressions.evaluate gives only the columns of the expression to pd.eval. Before, DataFrame.eval copied the index on each call. This copy used 90–95 % of the run time.
• estimate finds the inputs of DataMinus in one pass over the histograms. Before, it did one search for each variable and process.
• On sm2018_binned_v6, the fake-factor measurement takes 28 min instead of 180 min. estimate takes 9 min instead of approximately 45 min.