build: Downgrade ORT to 1.19.2 until OpenVINO NPU EP issue is resolved

[rmccorm4]

Downgrade to working version (1.19.2 used in r24.12) for stable support of ONNX Runtime OpenVino Execution Provider until we address the issues with NPU support.

[yinggeh]

I am fine with downgrading ORT verison but need @mc-nv to confirm.

[mc-nv]

I"m afraid that in r24.12 we been using version 1.20.1:
https://github.com/triton-inference-server/server/blob/r24.12/build.py#L77
Container data:

$ docker run --rm -it nvcr.io/nvidia/tritonserver:24.12-py3 readelf -V /opt/tritonserver/backends/onnxruntime/libonnxruntime.so  | grep gnu.version_d -A3
Version definition section '.gnu.version_d' contains 2 entries:
 Addr: 0x00000000000058f0  Offset: 0x000058f0  Link: 4 (.dynstr)
  000000: Rev: 1  Flags: BASE  Index: 1  Cnt: 1  Name: libonnxruntime.so.1
  0x001c: Rev: 1  Flags: none  Index: 2  Cnt: 1  Name: VERS_1.20.1

Believe our fix for this issue was using of previous version of the models based on OpenVINO 2023.3
https://github.com/triton-inference-server/server/blob/r24.12/qa/common/gen_qa_model_repository#L58

[rmccorm4]

Believe our fix for this issue was using of previous version of the models based on OpenVINO 2023.3
https://github.com/triton-inference-server/server/blob/r24.12/qa/common/gen_qa_model_repository#L58

This is a runtime issue and not a model generation issue. It passed with the previous versions in build.py for r24.11 (confirmed with a new pipeline 22573279 for this PR as well that all ONNX jobs pass), but failed when we updated ORT to 1.20 in r24.12 as well.

I'm happy to keep/upgrade the ORT version if we can fix the OpenVino EP issues - but I don't think we should keep it broken until we have clear bandwidth or steps to fix it.