Speed up RNN-T greedy decoding with cuda graphs

This uses CUDA 12.3's conditional node support.

Signed-off-by: Daniel Galvez <[email protected]>

nemo/collections/asr/modules/rnnt_abstract.py

@@ -282,7 +282,7 @@ def batch_select_state(self, batch_states: List[torch.Tensor], idx: int) -> List
     def batch_replace_states_mask(
         cls, src_states: list[torch.Tensor], dst_states: list[torch.Tensor], mask: torch.Tensor,
     ):
-        """Replace states in dst_states with states from src_states using the mask"""
+        """Replace states in dst_states with states from src_states using the mask, in a way that does not synchronize with the CPU"""
         raise NotImplementedError()
 
     def batch_split_states(self, batch_states: list[torch.Tensor]) -> list[list[torch.Tensor]]: