### Description
ReferenceEvaluator can take any proto as an input. This PR extends the
support to ModelContainer introduced in PR #5685.

### Motivation and Context
This makes it easier to test.

---------

Signed-off-by: Xavier Dupre <xadupre@microsoft.com>
Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>

…through protobuf (onnx#5685)

### Description

This PR proposes a way to build a large model (> 2Gb) in memory.

### Motivation and Context

The API for external data requires to create the whole model and then to
call a function to move big initializer to external files. It creates
class `LargeModelContainer`. It holds a ModelProto with no big
initializers. The big initiliazers are in an additional dictionary with
no protobuf structure (so it can be higher than 2Gb).

As an example, the creation of a large model:

```python
X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None])
Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None])
graph = make_graph(
    [
        make_node("MatMul", ["X", "A"], ["XA"]),
        make_node("MatMul", ["XA", "B"], ["XB"]),
        make_node("MatMul", ["XB", "C"], ["Y"]),
    ],
    "mm",
    [X],
    [Y],
    [
        # first large tensor, only the type and shape are used,
        # the location must start with '#' and be unique.
        make_large_tensor_proto("#loc0", "A", TensorProto.FLOAT, (3, 3)),
        from_array(np.arange(9).astype(np.float32).reshape((-1, 3)), name="B"),
        # second large tensor, only the type and shape are used
        make_large_tensor_proto("#loc1", "C", TensorProto.FLOAT, (3, 3)),
    ],
)
onnx_model = make_model(graph)

# The second parameter is a dictionary mapping the locations (or unique names) to the numpy arrays,
# it could be easily extended to support torch tensor.
large_model = make_large_model(
    onnx_model.graph,
    {
        "#loc1": (np.arange(9) * 100).astype(np.float32).reshape((-1, 3)),
        "#loc2": (np.arange(9) + 10).astype(np.float32).reshape((-1, 3)),
    },
)
large_model.check_model()

```

---------

Signed-off-by: Xavier Dupre <xadupre@microsoft.com>
Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
Co-authored-by: G. Ramalingam <grama@microsoft.com>
Signed-off-by: Linsho Kaku <linsho@preferred.jp>

### Description
ReferenceEvaluator can take any proto as an input. This PR extends the
support to ModelContainer introduced in PR onnx#5685.

### Motivation and Context
This makes it easier to test.

---------

Signed-off-by: Xavier Dupre <xadupre@microsoft.com>
Co-authored-by: Justin Chu <justinchuby@users.noreply.github.com>
Signed-off-by: Linsho Kaku <linsho@preferred.jp>