fix(expansion): keep empty fields when IFS delimiters repeat

[luantaraschi]

Field splitting treated every IFS character the same way: end the current field, and keep it if it has content. That is right for the whitespace characters, which collapse into a single separator. The other ones each end a field, even when the field is empty, so a:b::c is four fields.

With IFS=: and var holding a:b::c:

	bash	before	after
set -- $var; echo $#	4	3	4
printf '[%s]' "$@"	[a][b][][c]	[a][b][c]	[a][b][][c]

The same collapse hit ${arr[@]}, $(...) output and for loops, so a colon-separated record with an empty column came out one column short and the rest shifted left. read splits elsewhere and was already right, which is probably why this went unnoticed.

ifs.yaml already had six cases for it, all marked known_failure. They pass now, so the marker is gone from all six.

The rule the splitter follows, each part checked against bash:

• IFS whitespace before the first field is dropped, and a run of it ends one field.
• A non-whitespace delimiter ends a field. Whitespace on either side of it belongs to the same run, so a : b with IFS=': ' is two fields. A second non-whitespace delimiter in the same run ends a second, empty field.
• A leading delimiter leaves an empty first field. A trailing one does not add an empty last field.

is_ifs_whitespace spells out the six characters instead of calling char::is_whitespace, because bash goes by isspace in the C locale. The vertical tab counts, which char::is_ascii_whitespace leaves out, and the non-breaking space does not, which char::is_whitespace counts. There is a test case for the first half of that.

Six new cases cover the parts that had no coverage: a run of three or more delimiters, leading against trailing, whitespace around a delimiter, the whitespace character set, a delimiter split across two expansions in one word ($a$b where a ends in : and b starts with one), and an array expansion whose fields each split on their own.

One thing this does not fix: literal text in a word is still split, so IFS=:; set -- a::b gives three fields where bash gives one. That is #295, and the two cases covering it stay marked as known failures.