Add a scale op, related instances and examples

[geyyer]

• Add element op
• Add instances
• Add example
• Add client example

[bartekxk]

I have good idea for this problem:

• Add struct SCALAR : public BaseTensorLayout
• Pass Tuple<SCALAR, SCALAR, SCALAR> as layouts for Ds
• Pass only one value with shape ={1}, strides={0} for each d0,d1,d2
• ⁠Add private scale0, scale1, scale2 members in ConvScale elementwise, add some device function SetScales(scale0, scale1, scale2) and run it in global function
• in kernel add some if constexpr(is_same_v<Ds_layout.AT(i), SCALAR>), in this case set scale0, scale1, scale2 members to values under d0, d1, d2 pointer and pass empty D tuple to gridwise gemm

Pros of this solution:

• Easier API for the user (he dont need to pass some weird strides)
• Less memory usage (only need to allocate sizeof(datatype) bytes for each d0, d1,d2
• Kernel dont need to load d0, d1, d2 for each C store
• It allows to use CScalarPerVector > 1