Fix data loss / assert when encoding tiles with empty resolution levels

[nico]

When a tile is narrow but the number of decomposition levels is large, some of its lower resolution levels can become empty. Neither 5/3 not 9/7 DWT handled zero-length inputs correctly:

• opj_dwt_encode_and_deinterleave_h_one_row (5/3, lossless): The if (even) branch correctly checks if (width > 1), but the else branch for the odd case special-cased width == 1 and then just had an else branch on that. This caused width == 0 to be handled incorrectly. This is similar to the if (even) branch just above, and also similar to inverse opj_idwt53_h with its if (len > 1). Fixes opj_compress sometimes loses data when encoding in lossless mode #1575.

• opj_dwt_encode_and_deinterleave_h_one_row_real (9/7, lossy): only returned early for width == 1, so for width == 0 it called opj_dwt_encode_1_real() with dn == sn == 0. This tripped the assert(dn + sn > 1) therein in debug builds. Return early for width <= 1, to handle empty tiles too.

Before this commit, reproduced with a 119x53 grayscale image: width 119 with -t 117 leaves a 2-pixel-wide right edge tile, and -n 5 makes that tile's three lowest resolution levels empty:

# lossless: cols 117-118 corrupted
opj_compress -i in.pgm -o out.jp2 -n 5 -t 117,53

# lossy: assert abort in debug builds
opj_compress -i in.pgm -o out.jp2 -n 5 -t 117,53 -I

[nico]

I have a roundtrip test for this too, but I'll put it in a separate PR in case the fix is easier to merge than the test for some reason.