[HiCache] Fix Mamba track-boundary bookkeeping under overlap scheduling (#29792)

Co-authored-by: Hanming Lu <hanminglu@meta.com>

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_mamba.py

@@ -110,9 +110,9 @@ class TestUnifiedMambaHiCache(UnifiedRadixTreeTestMixin, CustomTestCase):
                 "--hicache-write-policy",
                 "write_through",
                 "--hicache-io-backend",
-                "direct",
+                "kernel",
                 "--hicache-mem-layout",
-                "page_first_direct",
+                "page_first",
                 "--max-total-tokens",
                 "12000",
                 "--max-mamba-cache-size",
@@ -171,9 +171,9 @@ class TestUnifiedMambaHiCacheL3(AccuracyTwoPassMixin, CustomTestCase):
                 "--hicache-storage-prefetch-policy",
                 "wait_complete",
                 "--hicache-io-backend",
-                "direct",
+                "kernel",
                 "--hicache-mem-layout",
-                "page_first_direct",
+                "page_first",
                 "--hicache-storage-backend",
                 "file",
                 "--max-mamba-cache-size",