In some input shapes, the current reduce schedule will underutilize the GPU.
E.g., reduce [1, 128, 128, 3] , dims=[1, 2] will spawn 1 threadblock with 3 threads that each iterate over 128*128 elements.
This PR made two changes to optimize these cases:

1. Add resolve_decompose in the resolve logic of Reduce. This will force launch separate kernels for each reduce dimension, increasing concurrency.
2. In the default reduce schedule template, spawn multiple warps within the reduce dimensions, which then will communicate via shared memory or use atomics to perform the reduce.

Also added a resolve rule for AdaptivePoolChannelLast.