Replace torch.jit.instance with instance to assist torch.compile

Summary: torch.jit.instance is mainly needed to refine Generics, but in this case instance is sufficient, instance unlike torch.jit.instance is also playing nicely with torch.compile

Differential Revision: D53797660

torchmultimodal/modules/layers/transformer.py

@@ -381,7 +381,7 @@ def _cross_attention_block(
             # TODO: figure out caching for cross-attention
             use_cache=False,
         )
-        assert torch.jit.isinstance(
+        assert isinstance(
             output, Tensor
         ), "cross-attention output must be Tensor."
         attention_output = self.cross_attention_dropout(output)