Add offset correction for split filter

[mh-northlander]

fix #148

Fix the offset when the length of the morpheme is different from the actual length in the input text.

Since correctOffset method is only available for CharFilter and Tokenizer, we need to keep offset mapping information for sudachi_split filter somehow.
I chose to put it in MorphemeAttribute because it is morpheme related data.

With combined character, extended mode behaves badly (e.g. "㍍㍉" will be extended to "㍍", "㍉", not "メ", "ー", "ト", "ル”, "ミ", "リ").
In this case using the normalized form will be more natural, but we cannot calculate offsets for them (mapping between surface and normalized form is missing). So I chose to keep using surface, i.e. text before normalization.

Also change dictionary type for test from small to core, to test "㍿" which is not in the small dict.

Note that due to the correctOffset behavior of icu-normalizer, offsets for subsplits of ㍿ (株式 + 会社) is now [0,0] + [0,1].