index_select: Optimizing the kernel with reducing for-loops in TensorInfo OffsetCalculator

[majing921201]

Two reasons for the slow perf in index_select

1. We used static loops times 12
2. We used int64_t for offset index, PVC doesn't have long datatype instruction, so it takes about 30us for once offset calculation.

So we have following optimization in this pr:
1, aligned CUDA, using dynamic loop boundry
2, optimized offset calculator

#816

We got 2x perf improvement in index_select

---

                                               Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg      Self XPU    Self XPU %     XPU total  XPU time avg    # of Calls

                                 aten::index_select        17.34%       2.161ms        41.05%       5.115ms      85.257us      12.734ms       100.00%      12.734ms     212.237us            60