Restore BSD-3-Clause copyright notices in source files derived from imagecodecs

[cgohlke]

Several Cython source files and related files were derived from imagecodecs, which is distributed under the BSD-3-Clause license. The per-file copyright and license headers present in every original imagecodecs source file are absent here, which violates clause 1 of that license.

The derivation is not in dispute: for example, _eer.pyx states in its own module docstring:

"The decoder is a self-contained excerpt from imagecodecs' BSD-3 licensed imcd.c (Christoph Gohlke), vendored under 3rdparty/imcd_eer/."

The 3rdparty/imcd_eer/LICENSE and 3rdparty/imcd_lzw/LICENSE files correctly attribute the vendored C code. The same must be done for Cython sources.

Every code derived from an imagecodecs source file needs the full original copyright notice restored.

[kevinjohncutler]

Ah sorry I missed this! Great points.

Rather than restore the notices, I took the derived code out and wrote our own. The EER decoder is written against the bitstream layout that RELION's renderEER.cpp documents, and validated against real Falcon 4 frames from EMPIAR-10568; the LZW encoder against TIFF 6.0 section 13. 3rdparty/imcd_eer and 3rdparty/imcd_lzw are both gone. Same for the two .pxd files that really were copies: libjxl.pxd is rewritten from the libjxl 0.11.2 headers, declaring only the part we call, and libultrahdr.pxd wasn't used by anything, so I deleted it.

They came out faster, which I honestly didn't expect going in. EER decode is 1.67x to 1.76x eer_decode across the three super-resolution levels on a real 4096x4096 frame (1.15 ms vs 2.02 ms at superres 0). LZW encode is 1.44x lzw_encode on 4 MB inputs, 188 vs 131 MB/s, with output within 0.03% of the same size in every case. The LZW win is mostly one trick: the dictionary is an open-addressed table whose slots carry the epoch they were written in, so CLEAR is epoch++ instead of a 32 KiB memset.

Since I'd clearly missed things by hand, I wrote a script that diffs every file we ship against your repo, by content as well as by name. It found two more: ricecomp.h, which was byte-identical to yours, and bcdec_dds.h, which nothing referenced. Both handled. What still matches your tree is the code we both vendor from the same upstream, like bitshuffle and qoi.h.

I did keep your fixes in rgbe.c and ricecomp.c, credited in the file headers. The rgbe overflow and partial-line fixes are the reason that file is safe on untrusted input, and the EOB error codes put back the checking cfitsio dropped in 3.08. I'd rather keep those than write something worse.

The rest of the .pyx files only mention imagecodecs in comments about matching your API or benchmarking against it. I don't think any of those are derived, but if you disagree I'm glad to go through them.

The repo also had no LICENSE at all, now fixed along with a THIRD-PARTY.md covering the vendored code and what the wheels bundle.