Use graph optimizer for gpu tensor prepack

[chenfucn]

Description

Use graph optimizer for cuda operator prepacking.

Motivation and Context

Our first sm80 quantized gemm kernel requires prepacking. Using current prepack infrastructure in gpu operator causes memory bloat -- old memory buffer not released.

So we define prepacking logic in a new graph optimizer. This solves the memory bloat problem. However, it also introduces the following problems:

1. Rewriting of the initializer tensors is restricted by operator shape inferencing rules. For example, there are 3 initializers for MatMulNBits, we can't combine them into a single tensor even if we want to. And we have to make sure the operator's shape inference logic does NOT verify the initializer's shape.
2. These rewriting logic is tightly coupled to each GPU operators. It really should be defined together with those operators, instead of defining them in a graph optimizer module.
3. The logic of prepacking depends on underlying GPU hardware. Currently this part is hard-coded for SM80.
4. In the operator, we have to expose a property to remember whether the operator is prepacked by the optimizer. This property is exposed to the user. But it should never be set by user.

[wejoncy]

I assume this can be supported by sm86, sm89?

[chenfucn]

it should

[wejoncy]

Would you support case=128? which is used widely.

[chenfucn]

it's difficult for this kernel. I am working on another version which hopefully can support that.

[chenfucn]

Restarting in another PR