Fix cache orientation: keep NeMo native [L, B, ...] throughout instead of transposing to batch-major (#1290)

config.json

@@ -18,14 +18,14 @@
   "decoder_hidden": 640,
   "decoder_layers": 1,
   "cache_last_channel_shape": [
-    1,
     17,
+    1,
     70,
     512
   ],
   "cache_last_time_shape": [
-    1,
     17,
+    1,
     512,
     8
   ]

streaming_encoder.int8.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ebc7619baf07e83dcd32072880c4a2144bb1713ef891a9abff81ff6a92422724
-size 133647770
+oid sha256:ceb642e54a352d5a7354102c0d5eba39db5f6b3936d36f3bf2819c26a5d0733a
+size 133645207