fold U+0345 after canonical ordering in utf8proc_decompose_custom - #379

utf8proc_decompose_custom case-folds each codepoint before the canonical ordering, but U+0345 (ccc 240) folds to U+03B9, a starter, so marks after it end up on the iota instead of the base letter and canonically equivalent strings fold differently: utf8proc_NFKC_Casefold maps U+1FB3 U+0301 to U+03B1 U+03AF but its NFC form U+1FB4 to U+03AC U+03B9 (fixes #3). this leaves U+0345 unfolded, also inside the 63 precomposed letters that decompose to it, and folds it once its run of marks ends at a starter of the canonical decomposition (dropped ignorables count), so the output matches the UCD's NFC(NFKC_CF(NFD(x))) on every codepoint and on 750k random strings and is byte-identical to before for input without U+0345. CHARBOUND is left as it was since its sentinels block the reordering, and there's a regression test in test/misc.c.