Update mktrace from gfit to Astropy fitting

[sean-lockwood]

Version 3.0.0. Builds onto PR #37.

• Remove dependence on deprecated stsci.tools.gfit Gaussian fitting; replace with Astropy fitting.
• Allow for scipy API change for scipy.signal.boxcar to scipy.signal.windows.boxcar.
• Changed Gaussian fitting initial guess to peak.
• Aggregate Astropy fitting warnings (with ierr > 4 into a single warning for each call of Trace.gFitTrace().
• Set data types to Float as they are created rather than after, avoiding conversions.
• Update test stdout. Will still require updating files (after manual validation) on Artifactory.

[sean-lockwood]

Early checks show differences in _crj_1dt.fits trace of 2e-3 to 5e-7 pixels, depending on the dataset.

It appears that gfit was more robust to initial conditions in the mean. So while setting the Astropy Gaussian model's mean to the peak location seems to work, this could result in different answers from gfit's center-of-range initial value.

We should probably construct a test suite across flux levels and modes in order to compare results before merging this change.

[sean-lockwood]

Validation work is ongoing. I compiled a test suite and have been measuring the scale of the resulting changes.

[sean-lockwood]

Some rows are not fitting the same regardless of changes to model initial parameters, e.g.:

[sean-lockwood]

Applying a Hann window to the fitting weights allows us to bias the result toward the center:

Going across all the columns of test data set o5f301020_crj.fits shows more consistent behavior:

[sean-lockwood]

Hann window weighting:

[stscirij]

Could this cause problems for E1 data?

[sean-lockwood]

I think it should be agnostic to Y-position unless it falls off the detector/subarray dimensions.

I did find some bad cases caused by the weights on the edges being zero, so changed the windowing slightly in the next commit.

[sean-lockwood]

Oh, I think I understand...

The window it's fitting each column in is 41 pixels tall (as set by the subim_size = 40 parameter).

[sean-lockwood]

Trace comparison: trace_results_5.pdf

Quick view of data for reference:
(Ignore the echelle datasets for now; I'll remove these soon.)

[sean-lockwood]

@stscirij -
I'm planning to present this to the STIS team after I'm satisfied with tuning the algorithm and validation testing.

The current code produces 4 output FITS files with 1D images. I'm considering merging these into a binary FITS table of a single product. Do you know of any dependencies on mktrace products that might break?