[wrapq] Operation Decomposition After Quantization

[dayo09]

What?

Problem Statement
When TICO converts (lower/decompose) quantized operator into several operators, some intermediate compute operations remain as float32.
For example,

Conv3d can be lowered into Conv2d + Add + Reshape... or Conv2d + Reshape.
NOTE that our Qwen3-VL patch embedding aims to be lowered down to Conv2d + Reshape without Add when patch size == stride size. This pass is scheduled to be implemented soon.
However, our current pass (26-02-13, main branch) converts Conv3d into Conv2d + Add + Reshape...

Thus, after quantization, Add is additionally generated. This extra Add remains in float32.

How to resolve?

1. Operator conversion should be done before quantization-calibration, if it produces any computing operations.
2. If required, we could manipulate qparams for specific operators.

[mhs4670go]

"remain as float32" means by that we can't calculate scale parameters of decomposed nodes except last node.

So, if such decomposition is needed, we should do it in model codes, which means the decompositon should be done before calibration.

The other solution could be re-calibration of whole graph in a export IR level instead of partial calibration.