Change input scale to double type for conv params. (#38346)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38346
Given qtensor stores scale as double, this mismatch can cause use to
repack weights everytime in QNNPACK. Worse given that we release
original weights runtime can crash.
Test Plan:
pytest test/quantization/test_quantized_module.py::TestStaticQuantizedModule::test_conv2d_api
Imported from OSS
Differential Revision: D21529384
fbshipit-source-id: 859b763dee5476e1554ebc278c5b95199a298eab