[FEA] support float16 dtype for many functions

[grlee77]

Description

We can easily preserve float16 dtype for many cucim.skimage.filters and cucim.skimage.transform functions because the underlying CuPy-based kernels already support it. This may generally be desirable, but two potential downsides are:

• it is a deviation from upstream scikit-image behavior which always promotes float16->float32
• a subset of algorithms/default parameters (e.g. default tolerance on iterative algorithms) may not be suitable for float16. In such cases we could still promote to larger dtypes as needed.

I think as long as we make the deviation from skimage dtype handling clear in the function docstrings it should be okay to allow operations on float16 images.