Update NaN block size tests for threshold_local function

[GenevieveBuckley]

Closes #256

This PR:

1. Removes some unnecessary tests. I wrote the tests several years ago, and I don't think there was a good reason to have them in there. It looks like I just wanted to have higher test code coverage at the time
2. It also explicitly specifies the like= kwarg when computing sigma from block_size for the gaussian local threshold function. I've found an edge case where da.asarray(cupy_array) correctly creates a dask array with cupy chunks, but da.asarray(cupy.nan) doesn't make a cupy-backed dask array (because type(cupy.nan) returns float). To avoid this problem, we use the like= kwarg to match the array type to the input image type.