fix: better handling for different operand dtypes in tf einsum

ivy/functional/backends/tensorflow/statistical.py

@@ -4,7 +4,6 @@
 
 # local
 import ivy
-from ivy.functional.ivy.statistical import _get_promoted_type_of_operands
 from ivy.func_wrapper import with_unsupported_dtypes
 from . import backend_version
 from ivy.utils.einsum_parser import legalise_einsum_expr
@@ -217,6 +216,15 @@ def einsum(
     *operands: Union[tf.Tensor, tf.Variable],
     out: Optional[Union[tf.Tensor, tf.Variable]] = None,
 ) -> Union[tf.Tensor, tf.Variable]:
-    dtype = _get_promoted_type_of_operands(operands)
     equation = legalise_einsum_expr(*[equation, *operands])
-    return tf.cast(tf.einsum(equation, *operands), dtype)
+    dtype_list = set(map(lambda x: x.dtype, operands))
+    dtype = dtype_list.pop()
+    if len(dtype_list) > 0:
+        for d in dtype_list:
+            dtype = ivy.promote_types(dtype, d)
+        dtype = ivy.as_native_dtype(dtype)
+        operands = list(
+            map(lambda x: tf.cast(x, dtype) if x.dtype != dtype else x, operands)
+        )
+
+    return tf.einsum(equation, *operands)