NEON-optimize 5-3 IDWT

[nico]

Takes bin/bench_dwt from 1.618 s to 0.432 s on my system.

---

Sadly, despite being a much bigger win in bench_dwt time, it's a much smaller end-to-end time improvement than #1629, as time spent in idwt is smaller than time spent in T1 / MQC decoding for lossless files (as there's more data). But still, it seems nice to have fast 5-3 IDWT anyways.

For a large lossless file I have, it takes bin/opj_decompress -i balloon-reversible.jp2 -o test.ppm -threads 0 from printing "decode time: 588 ms" to "decode time: 559 ms", around a 5% speedup for decoding. It's not nothing, but much less than the other PR.

I created that input file with bin/opj_compress -i image.ppm -o balloon-reversible.jp2 -M 1, where image.ppm is balloon.jp2 in decompressed (and balloon.jp2 is the usual jp2 balloon test file).

[nico]

For this one, I also verified that ctest --parallel has the same 50 failures it was without the PR, and I checked that dwt.c compiles fine with --target=armv7a-linux-gnueabihf --sysroot ~/src/chrome/src/build/linux/debian_bullseye_armhf-sysroot (to test 32-bit arm).