Fix a bug in large pickles

[yurymalkov]

Overflow caused by using uint32 in pointer operation. Large indexes are not saved properly.
commit changes uint32->uint64 and now the test is passing.

Thanks to Kai Wohlfahrt (https://www.linkedin.com/in/kwohlfahrt/ ) for reporting

Script to reproduce

import pickle
import numpy as np
import hnswlib

n = 5_000_000
d = 400


data=np.float32(np.random.random((n, 4)))
data=np.pad(data, ((0,0), (0, d-4)))
print(data.shape)

ann = hnswlib.Index("l2", d)
ann.init_index(n, ef_construction=10, M=16)

ann.add_items(data)
total_elements_to_test=n
print("testing recall")
labels, _ = ann.knn_query(data[:total_elements_to_test], k = 1)
print("recall:",np.sum(labels.reshape(-1)==np.arange(0,total_elements_to_test))/total_elements_to_test)


pickled=pickle.dumps(ann)
ann=pickle.loads(pickled)


print("testing recall after pickle")
labels, _ = ann.knn_query(data[:total_elements_to_test], k = 1)
print("recall:",np.sum(labels.reshape(-1)==np.arange(0,total_elements_to_test))/total_elements_to_test)
print(f"size whouls be at least{data.size * data.itemsize:,} bytes")
print(f"pickle size:{len(pickled):,} bytes")```