Add QuantizeLinear and DequantizeLinear for mixed precision

[kpu]

The current proposal has support for quantized types like tensor-quant8-asymm and some operators support them. Many networks run in mixed precision i.e. quantized output matrix multiply followed by logsoftmax in float32.

Propose adding https://github.com/onnx/onnx/blob/master/docs/Operators.md#DequantizeLinear and https://github.com/onnx/onnx/blob/master/docs/Operators.md#QuantizeLinear to make the quantized operators actually usable for many models.

[fdwr]

(4 years later 😲) quantize and dequantizeLinear are proposed here: #375 (comment). There's still some thought needed for the block size, given DequantizeLinear-21's new attribute (whether to do something similar/different/more generic/more limited...), but it has momentum.

Also related:

• WebNN should support int8 quantized models #128 (comment)
• WebNN should support NPU and QDQ operations #623
• Support block-wise quantization #779