fix: use torch.from_numpy() instead of torch.Tensor() to keep data type

[lsh918]

When running pytorch runner with async_run method in production mode(not in dev mode), tensors with other type than Float32 type is converted into Float32 type.(Ex. LongTensor is converted to FloatTensor) And I think this bug is originated by this line.

I made small test to check this issue. (numpy == 1.22.3 / torch == 1.12.1)

import torch
import numpy as np

arr = np.array([1, 2, 3, 4], dtype=np.int64)
t1 = torch.from_numpy(arr)
t2 = torch.Tensor(arr)

assert t1.dtype == torch.int64
assert t2.dtype == torch.float32

Thus I think changing torch.Tensor(ret) to torch.from_numpy(ret) should resolve this bug. Please tell me if it was a intended behavior.(i.e converting every tensor into Float32 tensor). Thank you!

[codecov]

Codecov Report

Merging #2951 (53ab5f0) into main (83d8a30) will decrease coverage by 1.62%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##             main    #2951      +/-   ##
==========================================
- Coverage   70.60%   68.97%   -1.63%     
==========================================
  Files         104      122      +18     
  Lines        9538    10135     +597     
==========================================
+ Hits         6734     6991     +257     
- Misses       2804     3144     +340     

Impacted Files	Coverage Δ
bentoml/_internal/frameworks/common/pytorch.py	72.63% <0.00%> (+72.63%)	⬆️
bentoml/picklable_model.py	0.00% <0.00%> (ø)
bentoml/serve.py	0.00% <0.00%> (ø)
bentoml/pytorch.py	0.00% <0.00%> (ø)
bentoml/bentos.py	49.33% <0.00%> (ø)
bentoml/sklearn.py	0.00% <0.00%> (ø)
bentoml/start.py	0.00% <0.00%> (ø)
bentoml/torchscript.py	55.55% <0.00%> (ø)
bentoml/keras.py	0.00% <0.00%> (ø)
bentoml/tensorflow.py	0.00% <0.00%> (ø)
... and 15 more

[larme]

LGTM, Thanks @lsh918 !