I don't think it makes sense to use the standard tokenizer with ICU. Instead, I recommend the icu_tokenizer
Also, as a side node, you should always apply stop word filter first, before lowercase or folding.
If the ICU plugin by Elastic really does not work, which would be very strange, I can offer an alternative implementation at https://github.com/jprante/elasticsearch-plugin-bundle/ where I just added an ICU folding filter test that succeeds like you have described https://github.com/jprante/elasticsearch-plugin-bundle/blob/403480349d5caf055e835c0dc3cf8bc798ec9359/src/test/java/org/xbib/elasticsearch/index/analysis/icu/IcuFoldingFilterTests.java