[Caching] Deterministic hashing of torch tensors

[lhoestq]

Currently this fails

import torch
from datasets.fingerprint import Hasher

t = torch.tensor([1.])

def func(x):
    return t + x

hash1 = Hasher.hash(func)
t = torch.tensor([1.])
hash2 = Hasher.hash(func)
assert hash1 == hash2

Also as noticed in https://discuss.huggingface.co/t/dataset-cant-cache-models-outputs/24945, using a model in a map function doesn't work well with caching. Indeed the bert-base-uncased model has a different hash every time you reload it. Supporting torch tensors may also help in this case.

This can be fixed by registering a custom pickling functions for torch tensors - as we did for other objects such as CodeType, FunctionType and Regex in py_utils.py