[Codegen] Add arith.clip/arith.clamp

[Groverkss]

Clipping or clamping is defined as:

clip(x, min_value, max_value) = min(max(x, min_value), max_value)

Some backends can generate better instructions if it's known we are clamping a value. For example, on gfx9, we have med3 instructions, which can be used to clamp a value in a range. See:https://github.com/ROCm/composable_kernel/pull/1182/files for an example of such optimization. Not only is it more efficient, it also gives better range information for a range analysis.

Tasks:

• Add arith.clip to upstream mlir (to arith/math, depending on discussion)
• Add optimized codegen to intrinsic for gfx9 for arith.clip

[elhewaty]

I will work on this. So the task is simply:

• Add the clip to the arith dialect at the llvm repo.
• when lowering arith dialect to LLVM, I should handle gfx9 case to use med3, right? are there any other cases that uses med3, or other better lowering?

[elhewaty]

@Groverkss @kuhar @ScottTodd, Ping