[ML] Update Deberta tokenizer (#116358)

* Was using byte position for end of offset, but it seems like using char position is correct

* Update docs/changelog/116358.yaml

* Update UnigramTokenizer.java

---------

Co-authored-by: Elastic Machine <elasticmachine@users.noreply.github.com>

docs/changelog/116358.yaml

@@ -0,0 +1,5 @@
+pr: 116358
+summary: Update Deberta tokenizer
+area: Machine Learning
+type: bug
+issues: []

x-pack/plugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/nlp/tokenizers/UnigramTokenizer.java

@@ -367,8 +367,10 @@ public final class UnigramTokenizer extends Tokenizer {
                         new DelimitedToken.Encoded(
                             Strings.format("<0x%02X>", bytes[i]),
                             pieces[i],
+                            // even though we are changing the number of characters in the output, we don't
+                            // need to change the offsets. The offsets refer to the input characters
                             offsetCorrection.apply(node.startsAtCharPos),
-                            offsetCorrection.apply(startsAtBytes + i)
+                            offsetCorrection.apply(endsAtChars)
                         )
                     );
                 }