modify the quantize.py file for efficiency

[yaozhewei]

1. Update the calculate from torch.split to torch.amin/torch.amax for fast computation
2. Update stochastic rounding computation logic (faster and cleaner) a. support both sym/asym sr in pytorch level b. reduce the new tensor creator from 2-->1 c. support cpu tensor as well
3. change fp16 --> fp32 to avoid overflow issue
4. change some other logic for easy understanding

[rocm-mici]

Can one of the admins verify this patch?

[mrwyattii]

Stale PR. quantize.py is quite different now. These changes are no longer relevant, therefore closing the PR.