KoreanTokenizer: TestRandomChains fails with incorrect offsets [LUCENE-10359]

It looks like KoreanTokenizer is causing this (NORI), but Kuromoji may be affected in the same way:

```
org.apache.lucene.analysis.tests.TestRandomChains > test suite's output saved to C:\Users\Uwe Schindler\Projects\lucene\lucene\lucene\analysis\integration.tests\build\test-results\test\outputs\OUTPUT-org.apache.lucene.analysis.tests.TestRandomChains.txt, copied below:
2> stage 0: e<[2-3] +1> ek<[4-6] +1> oy<[8-10] +1> 1<[11-12] +1> zzkuxp<[13-19] +1>
2> stage 1: e<[2-3] +1> ek<[4-6] +1> oy<[8-10] +1> 1<[11-12] +1> zzkuxp<[13-19] +1>
2> stage 2: e<[2-3] +1> e ek<[2-6] +0> ek<[4-6] +1> ek oy<[4-10] +0> oy<[8-10] +1> oy 1<[8-12] +0> 1<[11-12] +1> 1 zzkuxp<[11-19] +0>
2> stage 3: e<[2-3] +1> e ek<[2-6] +0> ek<[4-6] +1> ek oy<[4-10] +0> oy<[8-10] +1> oy 1<[8-12] +0> 1<[11-12] +1> 1 zzkuxp<[11-19] +0>
2> last stage: e<[2-3] +1> e ek<[2-6] +0> ek<[4-6] +1> ek oy<[4-10] +0> oy<[8-10] +1> oy 1<[8-12] +0> 1 zzkuxp<[11-19] +0>
2> TEST FAIL: useCharFilter=false text='?.e|ek|]oy{1 zzkuxp ZyzzV ycuqjnv axtpppvk \u233b\u23c8\u2314\u232e\u236e\u238d\u235e x d \"

'
2> Exception from random analyzer:
2> charfilters=
2> org.apache.lucene.analysis.pattern.PatternReplaceCharFilter(a, ifywufhi, java.io.StringReader@48586999)
2> org.apache.lucene.analysis.charfilter.MappingCharFilter(org.apache.lucene.analysis.charfilter.NormalizeCharMap@65036838, org.apache.lucene.analysis.pattern.PatternReplaceCharFilter@11d4ba35)
2> tokenizer=
2> org.apache.lucene.analysis.ko.KoreanTokenizer()
2> filters=
2> org.apache.lucene.analysis.en.KStemFilter(ValidatingTokenFilter@595d7938 term=,bytes=[],startOffset=0,endOffset=0,positionIncrement=1,positionLength=1,type=word,termFrequency=1,posType=null,leftPOS=null,rightPOS=null,morphemes=null,reading=null,keyword=false)
2> org.apache.lucene.analysis.shingle.ShingleFilter(ValidatingTokenFilter@13d08b48 term=,bytes=[],startOffset=0,endOffset=0,positionIncrement=1,positionLength=1,type=word,termFrequency=1,posType=null,leftPOS=null,rightPOS=null,morphemes=null,reading=null,keyword=false, u)
2> org.apache.lucene.analysis.util.ElisionFilter(ValidatingTokenFilter@6396b917 term=,bytes=[],startOffset=0,endOffset=0,positionIncrement=1,positionLength=1,type=word,termFrequency=1,posType=null,leftPOS=null,rightPOS=null,morphemes=null,reading=null,keyword=false, [fh, hiiwwxyyd, fcpodqor, qogvhmywr, l, icad])
2> Conditional:org.apache.lucene.analysis.ko.KoreanNumberFilter(OneTimeWrapper@5f0558f6 term=,bytes=[],startOffset=0,endOffset=0,positionIncrement=1,positionLength=1,type=word,termFrequency=1,posType=null,leftPOS=null,rightPOS=null,morphemes=null,reading=null,keyword=false)
> java.lang.IllegalStateException: last stage: inconsistent startOffset at pos=2: 8 vs 11; token=1 zzkuxp
> at __randomizedtesting.SeedInfo.seed([E4552C7844FC2DA3:8E0E93691DB20D50]:0)
> at org.apache.lucene.test_framework@10.0.0-SNAPSHOT/org.apache.lucene.tests.analysis.ValidatingTokenFilter.incrementToken(ValidatingTokenFilter.java:138)
> at org.apache.lucene.test_framework@10.0.0-SNAPSHOT/org.apache.lucene.tests.analysis.BaseTokenStreamTestCase.checkAnalysisConsistency(BaseTokenStreamTestCase.java:1130)
> at org.apache.lucene.test_framework@10.0.0-SNAPSHOT/org.apache.lucene.tests.analysis.BaseTokenStreamTestCase.checkRandomData(BaseTokenStreamTestCase.java:1028)
> at org.apache.lucene.test_framework@10.0.0-SNAPSHOT/org.apache.lucene.tests.analysis.BaseTokenStreamTestCase.checkRandomData(BaseTokenStreamTestCase.java:922)
> at org.apache.lucene.analysis.tests@10.0.0-SNAPSHOT/org.apache.lucene.analysis.tests.TestRandomChains.testRandomChainsWithLargeStrings(TestRandomChains.java:943)
```

---
Migrated from [LUCENE-10359](https://issues.apache.org/jira/browse/LUCENE-10359) by Uwe Schindler (@uschindler), updated Apr 09 2022

Start with TestRandomChains.java at line 943 and reproduce the failure involving KoreanTokenizer, then inspect the reported offset validation in ValidatingTokenFilter. Compare the stages in the saved output and check whether Kuromoji shows the same behavior. Done means the random-chain test no longer reports inconsistent startOffset values for this analyzer chain.

Written by the indexing model from the issue text.