WIP: Work with other float dtypes (float64, float16)

[hunse]

Just a start at this, since it came up when working on another repo.

TODO:

• Test that dtypes are properly conserved in layers and cells (i.e. output dtype matches layer dtype).
• Run some tests with multiple dtypes.
• Change epsilon (1e-8) used in e.g. Lowpass to work with other dtypes (e.g. the current value of 1e-8 will go to 0 if converted to float16). In the specific case of Lowpass, this epsilon probably doesn't need to be so small, since even np.exp(-1 / 1e-3) == 0.0.