Partially lower fusion to optimize deserialization performance.

[rdspring1]

This PR optimizes the deserialization time by only running the analyze step in GpuLower. We skip all the passes while retaining the necessary information to run a kernel. Deserialization is estimated to be ~75% faster than recompiling GpuLower completely.

After kernel compilation, we only need the information in GpuLower::KernelSummary to create a new ExecutorEntry at runtime. The kir::Allocate nodes are generated directly without the rest of the lowering process.

Our approach for creating kir::Allocate nodes is based on the ExpressionEvaluator. During serialization, we store a set of base IterDomains and a series of operations to create the TensorViews domains. The kir::Allocatenodes are not inserted into thekir::Kernel` because they are only used to calculate the size of buffers at kernel runtime.

TODOs:

• ExpressionSerializer and ExpressionBuilder builds TensorView and kir::Allocate nodes for global intermediate buffers and dynamic shared memory.
• Store all information to run fusion in KernelSummary, which is stored in FusionExecutor.
• Partially recreate and then deserialize KernelSummary
• Add support for LoadStoreOp
• Create separate paths for RNGOps and TMA kir nodes

[rdspring1]

!build

[rdspring1]

Need to update for new fusion executor dispatch system.