Use json lib to organize type shape during profiling

[RandySheriffH]

Solve an issue when composing input/output type and shape in json format.
The root cause of the issue is that when an op has tensor sequence or sparse tensor as input, the calculation of the number of inputs is wrong, which resulted in an extra comma appended to the end of a json array.
To solve the problem, we switch to a json lib for data organization and serialization.

[pranavsharma]

Just remembered that this will increase the binary size unconditionally. How much is the increase (please use the rel branch to calculate before and after)? If the increase is significant, may be for this release we can fix the issue without using this lib.

[RandySheriffH]

the lib is already introduced by session.lib, the dependency we added here actually reuse the binary no?

[pranavsharma]

The inclusion is protected by a macro:

onnxruntime/onnxruntime/core/session/inference_session_utils.h

Lines 8 to 16 in c716b56

	#if !defined(ORT_MINIMAL_BUILD)
	//
	// Includes to parse json session config from onnx model file
	//
	#include "core/graph/onnx_protobuf.h"
	#include "core/session/inference_session.h"
	#include "core/framework/session_options.h"
	#include "core/common/common.h"
	#include "nlohmann/json.hpp"

[RandySheriffH]

May be we should exclude it from minimal build as well.

[pranavsharma]

We shouldn't disable the whole profiling feature for minimal builds.

[RandySheriffH]

no, we just skip type and shape details.