Fix for xla_builder.Op.tuple return data type (#3231)
* [test_operations] add a test that checks data type equivalence
This change exposes a bug related with returned tensor data types,
with the newly added test "test_mixed_dtype_tuple" failing with
"AssertionError: torch.float32 != torch.bfloat16".
* Do not inherit logical_element_type in XLATensor::user_computation
... as suggested by @JackCaoG.
