floats with integer tensors not working with __r<op>__s

[Mgluhovskoi]

Came across this bug after rebasing recently and running the test cases from my PR #25. This bug results in allclose not being able to accept any integer tensors. When performing any __r<op>__ with a float scalar and an integer tensor you get the following error:

>>> b = tp.Tensor(1,tp.int32)
>>> b+1.1
==== Trace IR ====
t75 = storage(data=1, shape=(), dtype=int32, device=gpu:0)
t76 = storage(data=1.1, shape=(), dtype=int32, device=gpu:0)
t77 = t75 + t76
outputs:
    t77: [rank=(0), dtype=(int32), loc=(gpu:0)]

==== Flat IR ====
t75: [rank=(0), shape=(()), dtype=(int32), loc=(gpu:0)] = ConstantOp(data=1)
t76: [rank=(0), shape=(()), dtype=(int32), loc=(gpu:0)] = ConstantOp(data=1.1)
t_inter6: [rank=(1), shape=(()), dtype=(int32), loc=(gpu:0)] = ConstantOp(data=<mlir_tensorrt.runtime._mlir_libs._api.MemRefValue object at 0x74a500161c30>)
t_inter7: [rank=(1), shape=([0]), dtype=(int32), loc=(<class 'tripy.common.device.device'>)] = ConstantOp(data=<mlir_tensorrt.runtime._mlir_libs._api.MemRefValue object at 0x74a500160530>)
t_inter5: [rank=(1), shape=([0]), dtype=(bool), loc=(gpu:0)] = CompareOp.EQ(t_inter6, t_inter7, compare_direction='EQ')
t_inter8: [rank=(1), shape=(()), dtype=(int32), loc=(gpu:0)] = ConstantOp(data=<mlir_tensorrt.runtime._mlir_libs._api.MemRefValue object at 0x74a5001613b0>)
t_inter4: [rank=(1), shape=([0]), dtype=(int32), loc=(gpu:0)] = SelectOp(t_inter5, t_inter8, t_inter6)
t_inter3: [rank=(0), dtype=(int32), loc=(gpu:0)] = DynamicBroadcastOp(t75, t_inter4, broadcast_dim=[])
t_inter9: [rank=(0), dtype=(int32), loc=(gpu:0)] = DynamicBroadcastOp(t76, t_inter4, broadcast_dim=[])
t77: [rank=(0), dtype=(int32), loc=(gpu:0)] = AddOp(t_inter3, t_inter9)
outputs:
    t77: [rank=(0), dtype=(int32), loc=(gpu:0)]

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/tripy/tripy/frontend/tensor.py", line 219, in __repr__
    arr = self.eval()
  File "/tripy/tripy/frontend/tensor.py", line 183, in eval
    mlir = flat_ir.to_mlir()
  File "/tripy/tripy/flat_ir/flat_ir.py", line 153, in to_mlir
    map_error_to_user_code_and_raise(self, exc, stderr.decode())
  File "/tripy/tripy/backend/mlir/utils.py", line 458, in map_error_to_user_code_and_raise
    raise_error(
  File "/tripy/tripy/common/exception.py", line 195, in raise_error
    raise TripyException(msg) from None
tripy.common.exception.TripyException: 

--> <stdin>:1 in <module>()

TypeError: get(): incompatible function arguments. The following argument types are supported:
    1. (type: mlir_tensorrt.compiler._mlir_libs._mlir.ir.Type, value: int) -> mlir_tensorrt.compiler._mlir_libs._mlir.ir.IntegerAttr

Invoked with: IntegerType(i32), 1.1

Additional context:
Traceback (most recent call last):
  File "/tripy/tripy/flat_ir/flat_ir.py", line 145, in to_mlir
    mlir = to_mlir_impl()
  File "/tripy/tripy/flat_ir/flat_ir.py", line 84, in to_mlir_impl
    layer_outputs = op.to_mlir(layer_inputs)
  File "/tripy/tripy/flat_ir/ops/constant.py", line 89, in to_mlir
    attr = ir.DenseElementsAttr.get(attrs=to_attrs(self.data), type=self.outputs[0].to_mlir())
  File "/tripy/tripy/flat_ir/ops/constant.py", line 83, in to_attrs
    return [mlir_utils.get_mlir_scalar_attr(out_dtype, data)]
  File "/tripy/tripy/backend/mlir/utils.py", line 89, in get_mlir_scalar_attr
    return attr_func(get_mlir_dtype(dtype), value)
TypeError: get(): incompatible function arguments. The following argument types are supported:
    1. (type: mlir_tensorrt.compiler._mlir_libs._mlir.ir.Type, value: int) -> mlir_tensorrt.compiler._mlir_libs._mlir.ir.IntegerAttr

Invoked with: IntegerType(i32), 1.1

This occurs because of the following TraceIR: t76 = storage(data=1.1, shape=(), dtype=int32, device=gpu:0). This TraceIR is generated due to each __r<op>__ having the decorator: @frontend_utils.convert_inputs_to_tensors(sync_arg_types=[("self", "other")]) which will automatically create an int32 tensor from a float scalar eventually leading to an error.

Possible solutions:

1. If this type of operation is desired then we can upscale within __r<op>__ to float when scalar is a float and tensor is an int.
2. If this type of operation is not desired then can add some more checks that will throw helpful exceptions from __r<op>__ and will have to upscale within allclose since without this behavior allclose cannot handle any integer tensors.

[parthchadha]

The error is self explanatory now:

Refusing to automatically cast 'float32' argument to 'int32'.
    Hint: You may want to manually cast other operands in this expression to compatible types.
    Note: argument was: 

    --> <stdin>:1 in <module>()