Have a consistent concept/shared implementation of signal-to-noise

In the `snr` function and soon #509, we are using a definition of signal-to-noise that is a bit heuristic - basically whatever the uncertainty object is, we are assuming that's what the user meant as "noise".

But I think it may make more sense to instead assume the typical convention that "noise" means "gaussian statistics standard deviation" - i.e., we would want to try to convert the uncertainty object to StdDev and if it fails raise an exception asking the user to update the uncertainty to be meaningful.

A related task would be to have an underlying implementation of "what is the per-pixel signal to noise" that's part of the public API. The options that come to mind for me are either a `snr_pixel` function (`snr` is already taken for the "snr of a spectrum" operation), or a method on `Spectrum 1D` that would be `spectrum.pixel_snr` or similar.

Start with the existing `snr` function and the related work in #509, then inspect how uncertainty objects are currently interpreted. Compare the proposed `snr_pixel` function and `Spectrum 1D` method as public API options; done means a shared per-pixel signal-to-noise implementation with consistent uncertainty handling and a defined failure for unsuitable uncertainties.

Written by the indexing model from the issue text.