Hybrid conv kernel should use int8 weights instead of casted uint8.

Motivation for Hybrid conv to use int8:
- Currently the conv operation reinterpret_casts uint8 weights to int8 before using them in the kernel. This is an abuse of the schema's type system.
- We need support for shared weights in the QuantizeWeights tool. Currently if a shared weight is used by an operation that is consumed by a dequantize op and a hybrid op, we have no way to quantize the weights since one uses int8 and the other uses uint8. By specifying the type accurately for hybrid ops, we enable the tool to quantize shared weights. (We need to update the other hybrid kernels too.)

PiperOrigin-RevId: 226058081