fix(ai): update onnxruntime to 1.21.0+ for INT4 quantization support

The AI model build was failing to use INT4 quantization because
matmul_4bits_quantizer module requires onnxruntime 1.21.0 or higher.

Error:
  ImportError: cannot import name 'matmul_4bits_quantizer'
  from 'onnxruntime.quantization'

Issue:
- requirements.txt specified onnxruntime>=1.20.0
- matmul_4bits_quantizer module was added in a later release
- Build continued with FP32 (full precision) models instead of INT4
- Models were ~4x larger than necessary

Fix:
- Update requirements.txt to onnxruntime>=1.21.0
- Update build-sea.yml to match version requirement
- INT4 quantization will now work, reducing model sizes by ~75%

This provides better model compression without accuracy loss.

Author: jdalton

.github/workflows/build-sea.yml

@@ -313,7 +313,7 @@ jobs:
           if [ "${{ steps.ai-cache-valid.outputs.valid }}" != "true" ]; then
             echo "::group::Installing Python dependencies"
             python3 -m pip install --upgrade pip
-            python3 -m pip install transformers torch optimum[onnx] "onnxruntime>=1.20.0"
+            python3 -m pip install transformers torch optimum[onnx] "onnxruntime>=1.21.0"
             echo "::endgroup::"
           fi
 

packages/socketbin-cli-ai/requirements.txt

@@ -1,4 +1,4 @@
 transformers
 torch
 optimum[onnx]
-onnxruntime>=1.20.0
+onnxruntime>=1.21.0