Feature- calibration polynomials in the wavelength domain

Starfish currently supports calibration polynomials in the flux domain, parameterized as Chebyshev polynomials (c.f. Czekala et al. 2015). In principle there could be calibration uncertainties attributable to wavelength-domain instrumental artifacts-- basically bad wavelength solution in your data reduction pipeline. The v_z term takes up the zeroth order instrumental artifact, a wavelength shift (of course, in perfectly or near-perfectly calibrated spectra, v_z has astrophysical significance!). However linear or higher-order distortions of your wavelength solutions are possible. In fact, they're not only possible, they're omni-present, but just at a level that may be negligible for most applications.

In any case, if for some reason you are stuck with spectra with poor wavelength calibration, one could straightforwardly implement wavelength-domain calibration polynomials. Such a feature would need to modify the over-sampled wavelength axis in the update_theta() step. This change would add negligibly to the computation time-- the FFT would proceed as normal.

To be clear this is a feature idea enabled by this powerful spectral inference framework, not some inadequacy as might be misconstrued by the GitHub label "Issue", which takes on multiple meanings.

Start by locating the model's update_theta() step and tracing how the over-sampled wavelength axis is built and used by the FFT. Define how linear and higher-order wavelength-domain calibration polynomials should interact with v_z and existing flux-domain calibration. Done means wavelength-domain calibration is supported without disrupting the existing inference and FFT workflow.