Prevent complex<Half> accumulation overflow in sum_functor specialization

[yucai-intel]

Problem Description:
The XPU backend did not correctly enable the high-precision accumulator rule when processing the sum reduction of complexHalf, causing premature overflow of low-precision accumulators in the kernel's internal accumulation operation.

Specific Modifications and Effects:

• Modify the reduce_dispatch function: The original scheduler logic omitted explicit high-precision boosting for ComplexHalf. The fix forces the acc_t to be boosted to complex at the scheduling entry point, ensuring the correct precision is used for reduction.
• Modify the sum_functor specialization: The kernel's underlying implementation may incorrectly downgrade the accumulator type to the low-precision out_t. The fix forces the out_t parameter of gpu_reduce_kernel to be boosted to complex to avoid the type downgrading bug within the kernel.

[yucai-intel]

For #2008