[Build] Failed test XnnpackEP.TestQDQConvS8S8_per_channel

### Describe the issue

when build from a clean master checkout
```
./build.sh --parallel 8 --config Release --use_xnnpack
```
I get two failed xnnpack tests
```
1: [ FAILED ] XnnpackEP.TestQDQConvS8S8_per_channel
1: [ FAILED ] XnnpackEP.TestConvTranspose_qdq
```

### Urgency

_No response_

### Target platform

Linux x86_64

### Build script

`./build.sh --parallel 8 --config Release --use_xnnpack`

### Error / output

I get
```
1: [ RUN ] XnnpackEP.TestQDQConvS8S8_per_channel
1: 2023-02-03 12:22:34.985051001 [W:onnxruntime:xnnpack_qdq_test_graph_conv_s8s8_perchannel, session_state.cc:1136 VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which may or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
1: 2023-02-03 12:22:34.985087567 [W:onnxruntime:xnnpack_qdq_test_graph_conv_s8s8_perchannel, session_state.cc:1138 VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assignments.
1: /home/zhukov/clients/rbx/onnxruntime/onnxruntime/test/util/test_utils.cc:50: Failure
1: Value of: ltensor.DataAsSpan()
1: Expected: contains 21632 values, where each value and its corresponding value in 16-byte object <80-54 00-00 00-00 00-00 40-BA 2D-40 E2-7F 00-00> are an almost-equal pair
1: Actual: 16-byte object <80-54 00-00 00-00 00-00 80-E7 3C-40 E2-7F 00-00>, where the value pair (8.1733408, 6.53867292) at index #704 don't match, which is -1.63467 from 8.17334
1: Google Test trace:
1: /home/zhukov/clients/rbx/onnxruntime/onnxruntime/test/common/tensor_op_test_utils.cc:14: ORT test random seed: 2900955797
1: [ FAILED ] XnnpackEP.TestQDQConvS8S8_per_channel (15 ms)
```

and
```
1: [ RUN ] XnnpackEP.TestConvTranspose_qdq
1: 2023-02-03 12:22:35.071365265 [W:onnxruntime:test_conv_follow_convtrans_s8, session_state.cc:1136 VerifyEachNodeIsAssignedToAnEp] Some nodes were not assigned to the preferred execution providers which ma
y or may not have an negative impact on performance. e.g. ORT explicitly assigns shape related ops to CPU to improve perf.
1: 2023-02-03 12:22:35.071381179 [W:onnxruntime:test_conv_follow_convtrans_s8, session_state.cc:1138 VerifyEachNodeIsAssignedToAnEp] Rerunning with verbose output on a non-minimal build will show node assign
ments.
1: /home/zhukov/clients/rbx/onnxruntime/onnxruntime/test/util/test_utils.cc:50: Failure
1: Value of: ltensor.DataAsSpan()
1: Expected: contains 4968 values, where each value and its corresponding value in 16-byte object <68-13 00-00 00-00 00-00 40-69 2D-40 E2-7F 00-00> are an almost-equal pair
1: Actual: 16-byte object <68-13 00-00 00-00 00-00 80-48 3C-40 E2-7F 00-00>, where the value pair (3.6200695, 4.22341442) at index #0 don't match, which is 0.603345 from 3.62007
1: Google Test trace:
1: /home/zhukov/clients/rbx/onnxruntime/onnxruntime/test/common/tensor_op_test_utils.cc:14: ORT test random seed: 2900955797
1: [ FAILED ] XnnpackEP.TestConvTranspose_qdq (28 ms)
```

### Visual Studio Version

_No response_

### GCC / Compiler Version

gcc version 11.3.0 (Ubuntu 11.3.0-1ubuntu1~22.04)