[model] add least squares solver (review & merge)

[ryanhammonds]

Addresses #5. I made this because I was applying this package to 2d spectrograms and needed a faster solver. Allows directly solving conv weights as a linear system, rather than using iterative optimization. This should be faster than convolutional.

Additionally:

• a dilation kwarg was added that allows the expansion/dilation of the kernel to account for learn weights at various spatial scales, e.g. dilation=2 constructs a linear system from sliding the following kernel over an input:

• a equidistant kwarg was added that allows the removal of weight constraints when False. This is for cases when the symmetry assumption is violated.

To-Do Before Merging

• Tests
• Confirm predictive value of weights
• Ensure it's actually faster the convolution