Add (experimental) Nikon HE compression Support - #835

Nikon's headers differ from the standard in three spots, but only one
of them matters: every component carries a third byte saying how it was
decomposed. Plain JPEG XS assumes only the last component can skip the
wavelet, and Nikon's is component 2.

Bands are ordered by component, packets by subband, and the weights
table follows the packet order. Get that mapping wrong and nothing
complains, because the code is unary: wrong values still take up the
right number of bits, so the file parses cleanly and only the pixels
come out wrong. A test pins the mapping down.

Nikon always uses the long packet header, even where the standard's
rule of thumb would pick the short one, and it writes sign bits for
coefficients that run past the end of a short last group. Those have to
be skipped, or everything after them comes out with the wrong sign.

The tests compare against the C reference decoder; point
RAWLER_JPEGXS_REFERENCE at its output directory to run them.

A precinct holds two plane lines, but finishing the second one needs
data from the precinct after it, so the output runs one precinct behind
the input.

Each component says for itself whether it skips the wavelet, so the
code reads that rather than assuming it is the last one. Nikon's is
component 2.

The four components are a star-tetrix decorrelation of the Bayer planes
rather than the planes themselves, and the header never says so: Cpih
reads 0 and the markers that would carry the parameters aren't there.

Nikon also scales the chroma planes and square-root compands the
samples. Those constants are fitted from a single scene, so they're
approximate, and there's no telling whether they hold for other
cameras.

These files used to be rejected outright; now they decode, and the
result correlates 0.99 with a lossless shot of the same scene.

Only tested on a Z6III.