[WIP] Replace shift_group_left with reduce_over_group in Reduce Sum

[CuiYifeng]

This pull request refactors and enhances the reduction utilities in the SYCL backend to support native SYCL reduction operations when available, improving performance and maintainability. The most important changes are grouped below:

Support for Native SYCL Reduction Operations

• Introduced the get_native_sycl_op and native_sycl_op_t type traits to detect and propagate native SYCL reduction operators from functors, enabling the use of optimized native reductions when possible.
• Updated the SumFunctor to define a native_sycl_op (sycl::plus), allowing it to leverage SYCL’s built-in reduction.

Refactoring of Reduction Logic

• Added the tree_reduce function, which uses the native SYCL operator for floating-point types if available, and otherwise falls back to a manual reduction loop. This is now used in place of the previous manual reduction code in group_reduce and group_x_reduce.

Integration and Propagation of Native Operators

• Modified the func_wrapper_t and reduction kernel logic to propagate the detected native SYCL operator through the reduction pipeline, ensuring that the most efficient reduction method is chosen at compile time.

These changes collectively make the reduction code more generic, extensible, and able to utilize hardware-accelerated reductions where possible, while falling back to custom logic otherwise.