Deconvolution with provided output shape inconsistency between MNN and ONNXRuntime

[h6197627]

There is inconsistency between outputs produced by MNN and ONNXRuntime for Deconvolution op with manually specified output shape ONNX model (single ConvTranspose layer) MNN model (converted from ONNX model with 3.2.5 converter)

Compare script:

import numpy as np
import MNN # version 3.2.5
import onnxruntime as rt # version 1.23.2

if __name__ == '__main__':
	input_data = np.random.rand(2, 3, 8, 8).astype(np.float32)

	ort_model = rt.InferenceSession('deconv.onnx', providers=rt.get_available_providers())
	ort_output = ort_model.run(['output'], {'input' : input_data})[0]

	mnn_model = MNN.nn.load_module_from_file('deconv.mnn', ['input'], ['output'])
	mnn_input = MNN.expr.placeholder(input_data.shape, MNN.expr.NCHW)
	mnn_input.write(input_data)
	mnn_output = mnn_model.forward([mnn_input])[0].read()

	print(np.allclose(ort_output, mnn_output, rtol=1e-4, atol=1e-4))
	np.save('input.npy', input_data)
	np.save('output_ort.npy', ort_output)
	np.save('output_mnn.npy', mnn_output)

After some investigation it turned out that current MNN output is the same as ONNXRuntime before these changes: https://github.com/microsoft/onnxruntime/commit/6246662b1d1b0ba33a9eebfcc426397f62335f82 https://github.com/microsoft/onnxruntime/commit/f96f2225262ed9aaa17604aeb3185b98c5dc71d2

Also checked that OpenVINO output is the same as in current ONNXRuntime version.