[tokenizers] Fixes memory leak when there is overflowing tokens (#3317)

* If you call TokenizersLibrary.LIB.getOverflowing you must also clean up all overflow encodings.
If withOverflowingTokens was false no Encodings where generated leaving jni Encoding handles that would not be properly deleted.

This introduces a new native method where you can inquire about number of overflow tokens without using any jni resources.
And you will now only call TokenizersLibrary.LIB.getOverflowing(encoding) if withOverflowingTokens is true.

extensions/tokenizers/rust/src/lib.rs

@@ -497,6 +497,17 @@ pub extern "system" fn Java_ai_djl_huggingface_tokenizers_jni_TokenizersLibrary_
     array
 }
 
+#[no_mangle]
+pub extern "system" fn Java_ai_djl_huggingface_tokenizers_jni_TokenizersLibrary_getOverflowCount(
+    _: JNIEnv,
+    _: JObject,
+    handle: jlong,
+) -> jint {
+    let encoding = cast_handle::<Encoding>(handle);
+    let count = encoding.get_overflowing().len();
+    count as jint
+}
+
 #[no_mangle]
 pub extern "system" fn Java_ai_djl_huggingface_tokenizers_jni_TokenizersLibrary_getOverflowing<
     'local,

extensions/tokenizers/src/main/java/ai/djl/huggingface/tokenizers/HuggingFaceTokenizer.java

@@ -608,10 +608,11 @@ private Encoding toEncoding(long encoding, boolean withOverflowingTokens) {
         long[] specialTokenMask = TokenizersLibrary.LIB.getSpecialTokenMask(encoding);
         CharSpan[] charSpans = TokenizersLibrary.LIB.getTokenCharSpans(encoding);
 
-        long[] overflowingHandles = TokenizersLibrary.LIB.getOverflowing(encoding);
-        boolean exceedMaxLength = overflowingHandles.length > 0;
+        int overFlowCount = TokenizersLibrary.LIB.getOverflowCount(encoding);
+        boolean exceedMaxLength = overFlowCount > 0;
         Encoding[] overflowing;
         if (withOverflowingTokens) {
+            long[] overflowingHandles = TokenizersLibrary.LIB.getOverflowing(encoding);
             overflowing = new Encoding[overflowingHandles.length];
             for (int i = 0; i < overflowingHandles.length; ++i) {
                 overflowing[i] = toEncoding(overflowingHandles[i], true);

extensions/tokenizers/src/main/java/ai/djl/huggingface/tokenizers/jni/TokenizersLibrary.java

@@ -60,6 +60,8 @@ public native long[] batchEncodePair(
 
     public native long[] getOverflowing(long encoding);
 
+    public native int getOverflowCount(long encoding);
+
     public native String decode(long tokenizer, long[] ids, boolean addSpecialTokens);
 
     public native String getTruncationStrategy(long tokenizer);