[ENH] StatsModels: Add invalid_inputs parameter to allow dropping of missing data columns to run nodes

[coryshain]

BIDSStatsModelsNode.run() gracefully drops missing contrasts when invalid_contrasts=="drop", but missing data columns in the GLM at the run level (entries of Model.X) currently always throw an error. As far as I can tell, this makes it impossible to use BIDSStatsModelsNode (and BIDS apps that depend on it, like fitlins) for tasks in which some conditions do not appear in some runs (a use case I happen to have at the moment, which also seems to pop up for others, e.g., here and here). This PR adds an invalid_inputs parameter BIDSStatsModelsNode.run() and BIDSStatsModelsNodeOutput.__init__() that permits a "drop" value just like invalid_contrasts, which warns about missing data columns and then removes them from var_names. I've tested this modification with fitlins and it appears to give the desired enhancement: individual runs estimate all and only the available contrasts (and missing contrasts at downstream nodes are then handled by the existing invalid_contrasts feature). I've set the default value of invalid_inputs to "error", so that the previous default behavior remains unchanged.