Make signal functions respect input dtype.

[luigifcruz]

This pull request ensures that the data type from the input and output are the same. This results in an expressive 35% speedup in single-precision Hilbert operations.

## Upstream
%timeit sc.hilbert(data64)
126 ms ± 284 us per loop (mean ‡ std. dev. of 7 runs, 10 Loops each)
%timeit sc.hilbert(data32)
101 ms ± 662 us per loop (mean ‡ std. dev. of 7 runs, 10 Loops each)

## Patch
In [5]: %timeit sc.hilbert (data64)
130 ms ± 259 us per loop (mean + std. dev. of 7 runs, 10 Loops each)
In [6]: %timeit sc.hilbert(data32)
75.6 ms ± 174 us per Loop (mean ‡ std. dev. of 7 runs, 10 Loops each)

[luigifcruz]

I just noticed that the lfilter_zi function also doesn't respect the input type. The previous commit applies patches for this. A test for the function was also added. This commit might require re-review. cc @larsoner @ilayn

[luigifcruz]

The CI is now throwing three test errors after the last commit. I can't reproduce them in my local machine. I'm not familiar with RuntimeWarning: invalid value encountered in ? (vectorized) error. I believe this happens only in 32-bits systems.

[luigifcruz]

I tried to reproduce the errors reported by the CI with a 32-bits Debian VM but no luck. This error appears to be similar to this Satpy error. Any theories appreciated!

[larsoner]

A restart fixed the errors, this LGTM, @tylerjereddy's comment has been addressed, and @ilayn approved already, so in it goes. Thanks @luigifcruz !