Barycentric correction per-fiber

From David Schlegel: Thinking about the barycentric corrections again, the worst-case
scenario for errors by using the tile center instead of the fiber position is
(cos(45+1.6 deg) - cos(45 deg)) * 30 km/s = 0.6 km/s
Across a tile, that’s 1.2 km/s peak-to-peak error.

Sky-subtraction needs to be done on the as-observed wavelength
scale, meaning either before the barycentric correction or with the
same done per fiber.

But a way to fix this would be to make small offsets to the final resolution
matrix to account for the residual per-fiber errors. The central fibers
have no change, nor do fibers offset in the perpendicular direction to
the earth’s motion vector. Fibers parallel to that motion would have
their resolution matrices shifted by ~0.01 - 0.02 pix.

This approach would give us data products that are unchanged, with
the same wavelength scale for all fibers. But for the detailed analysis,
like projecting stellar models to the as-observed spectra, this would
formally have the correct velocity information.