Fix array_to_datum/datum_to_array crashing on NumPy >= 2.0

[Chessing234]

Bug

python/caffe/io.py's datum (de)serializers still use two NumPy entry points that were deprecated and then removed in NumPy 2.0:

# array_to_datum, line 76
if arr.dtype == np.uint8:
    datum.data = arr.tostring()

# datum_to_array, line 89
if len(datum.data):
    return np.fromstring(datum.data, dtype=np.uint8).reshape(
        datum.channels, datum.height, datum.width)

• numpy.ndarray.tostring was deprecated in NumPy 1.19 in favour of tobytes and removed in NumPy 2.0.
• numpy.fromstring for binary data was deprecated in NumPy 1.14 in favour of frombuffer and removed in NumPy 2.0.

On any user running NumPy ≥ 2.0, array_to_datum (save path) and datum_to_array (load path for uint8 datums, e.g. image datasets) raise AttributeError: 'numpy.ndarray' object has no attribute 'tostring' and AttributeError: module 'numpy' has no attribute 'fromstring' respectively. This breaks pycaffe data ingestion for anyone on a modern scientific-Python stack.

Fix

Switch to the binary-safe replacements NumPy's deprecation notices have been directing users to for years:

     if arr.dtype == np.uint8:
-        datum.data = arr.tostring()
+        datum.data = arr.tobytes()

     if len(datum.data):
-        return np.fromstring(datum.data, dtype=np.uint8).reshape(
+        return np.frombuffer(datum.data, dtype=np.uint8).reshape(
             datum.channels, datum.height, datum.width)

Both replacements are byte-for-byte equivalent to the deprecated calls: ndarray.tobytes() produces the same bytes as tostring() (the name was changed precisely because "tostring" misled users into expecting a Python str), and np.frombuffer(buf, dtype=np.uint8) reads the same bytes back into the same uint8 ndarray shape. Existing NumPy 1.x users are unaffected — both tobytes and frombuffer have been present since NumPy 1.9 and 1.14 respectively.

Two one-token renames, no other changes.