This PR is aim to reduce memory footprint of the base GWR model from method in FastGWR. Two approaches are 1) not to store N by N weight matrix 2) not to store N by N hat matrix.

For weight matrix, at every location, a row vector of weights is calculated but not stored. This may add some computation but very trivial. The entire N by N W matrix can be computed on demand as a method to GWRResults class.

For hat matrix, only tr_S and tr_STS are needed and they are calculated inside of GWR.fit() fitting loop additively without storing S. A boolean parameter hat_matrix is offered for whether to store the hat matrix. By default, hat_matrix is set to False for GWR. However, in MGWR, hat matrix of GWR is used to compute projection matrices of MGWR, so hat_matrix is switched to True behind the scene for calibrating MGWR.

Overall speed performance is similar, but memory footprint is much lower. Based on a dataset with 10k locations, peak memory drops from 4487MB to 155MB on my laptop.