fix bug in sq torch backend

[xin3he]

Type of Change

bug fix, related to latest torch modification.

Description

NLPTOOLKIU-861
Fake quantized model for smoothquant on torch backend cannot be reload correctly.

(Pdb) m = torch.nn.quantized.Quantize(1, 0, dtype=torch.uint8)    
(Pdb) m.scale.dtype                                                    
torch.int64  # cannot load float scale
# change it as below
(Pdb) m = torch.nn.quantized.Quantize(1.0, 0, dtype=torch.uint8)       
(Pdb) m.scale.dtype                 
torch.float32

Expected Behavior & Potential Risk

reload model accuracy recover

How has this PR been tested?

local tested