[Perf] improve the hash kernel for mm

[mickqian]

The current gpu_tensor_hash implementated in #5974 has following drawbacks:

1. add itself is not a very decent reduction method
2. will perform an on-cpu reduction, which is not very performant for large tensors

TODO

1. Rewrite a performant and robust tensor hash function
2. Test the performance, consistency and correctness of the hash function against real data

Reference

You can reference here for inspirations

[mickqian]

@yhyang201

[yhyang201]

There isn’t a standard solution for GPUs, so I’ll try using blake3 on the CPU to see if we can achieve further optimization.