variables: Add sparse/dense mode for to to_df, and add alternative design matrix API to access sparse variables

[adelavega]

pybids/bids/variables/collections.py

Lines 610 to 620 in f4bf70c

	# If all variables are sparse/simple, we can pass them as-is. Otherwise
	# we first force all variables to dense via .resample().
	if all(isinstance(v, SimpleVariable) for v in _vars):
	variables = _vars
	else:
	sampling_rate = sampling_rate or self.sampling_rate
	var_names = [v.name for v in _vars]
	collection = self.resample(
	sampling_rate, variables=var_names, force_dense=include_sparse
	)
	variables = list(collection.variables.values())

[adelavega]

to_df casts everything to dense, which is a problem for the HRF property of BSM.

To be able to hand off HRF convolution to another program we need to be able to output sparse variables seperately from dense variables.

We need to add a mode that separate dense/sparse, and a different way of accessing the design matrix (aside from .X).

[adelavega]

@jmumford also requests this in order to look at the unconvolved regressors for model inspection

[adelavega]

This is actually more complex than it looks, because BIDSStatsModelsNodeOutput would need to keep in parallel both dense and sparse parts of the DM, so all logic downstream (i.e. handling intercept, expanding wild cards, etc), would have to be refactor to handle keeping the DM split.