feat: NeMo-style tokenization, 455/470 (97%)

Replace tagger.star with NeMo-style token + closure(delete_extra_space
+ token) pattern. This ensures explicit space consumption between
tokens, resolving many segmentation ambiguities:
- seven eleven stores => 7-eleven stores (whitelist now wins)
- set alarm at ten to eleven pm => set alarm at 10:50 p.m.

Author: pengzhendong

itn/english/inverse_normalizer.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from importlib_resources import files
+from pynini import closure
 from pynini.lib.pynutil import add_weight, delete
 
 from itn.english.rules.cardinal import Cardinal
@@ -67,8 +68,9 @@ def build_tagger_and_verbalizer(self):
             | add_weight(char.tagger, 100)
         ).optimize()
 
-        tagger = tagger.star
-        self.tagger = tagger @ self.build_rule(delete(" "), "", "[EOS]")
+        token = tagger
+        graph = token + closure(self.DELETE_EXTRA_SPACE + token)
+        self.tagger = delete(" ").star + graph + delete(" ").star
 
         verbalizer = (
             cardinal.verbalizer

itn/english/test/data/en_telephone.txt

@@ -15,3 +15,4 @@ a b nine => ab9
 a b c => a b c
 five w k r a three one => 5wkra31
 x three eighty six => x386
+r t x forty fifty t i => RTX 4050ti

itn/english/test/data/en_time.txt

@@ -25,3 +25,4 @@ half past twelve => 12:30
 quarter past one => 01:15
 quarter to one => 12:45
 quarter to twelve => 11:45
+set alarm at ten to eleven pm => set alarm at 10:50 p.m.

itn/english/test/data/en_whitelist.txt

@@ -4,6 +4,8 @@ mister dao => mr. dao
 saint george => st. george
 i like for example ice cream => i like e.g. ice cream
 s and p five hundred => S&P 500
+seven eleven stores => 7-eleven stores
 r t x => RTX
 cat five e => CAT5e
 c u d n n => cuDNN
+p c i e x eight => PCIe x8