Temporarily pin onnxruntime to <1.16

[mguaypaq]

Due to microsoft/onnxruntime#17631, version 1.16.0 is causing CI to fail on master. We can change the pin to !=1.16.0 (or to however many versions are buggy) once upstream releases a fix.

Part of #4225.

[joshuacwnewton]

Reading through the underlying onnxruntime issue, it seems like this may also be easily fixed by specifying the CPU provider to the providers argument:

The workaround is to explicitly supply the CPUExecutionProvider in the list.

Since onnxruntime (the cpu version) is what we currently use, as opposed to onnxruntime-gpu, I think it should be fine to specify this explicitly?

In deepseg_/onnx.py:

-    ort_sess = ort.InferenceSession(model_path, sess_options=sess_options)
+    ort_sess = ort.InferenceSession(model_path, sess_options=sess_options, providers=['CPUExecutionProvider'])

[mguaypaq]

This is what's suggested by the error message, but I kind of agree with the explanation under Describe the issue that this may be misleading. The docs show the Azure provider as preview-only (link), and show a "Default CPU" provider, and onnxruntime.InferenceSession shows providers as being an optional argument (link), so I don't think it's meant to be mandatory.

[mguaypaq]

But ultimately, I'm fine with either fix, up to you.

[joshuacwnewton]

I think my thinking here was that we might want to keep the versions as open as possible? But, that said, I have no guarantee whether the providers fix is backwards-compatible with older verisons. So, on second thought, I think I like your fix a little more -- we can always go back and specify onnxruntime!=1.16.0 once the patch comes out.

[mguaypaq]

Looking at the changelog for ORT 1.9, which is mentioned in the error message. Looking at the first point under Packages, it looks like this was meant for the GPU packages only, not for the basic CPU package.