C Unpickler memory leak when unpickling big geojsons

Bug report

Bug description:

I noticed that the pickle.load method leaves more memory allocated than just the size of the returned object when unpickling geojson dictionaries.

I initially noticed the leak in my docker container running on x86. Here is a code snippet that reproduced the problem for me on M1 MacOS Sonoma (python 3.11.9).

import io
import pickle

import psutil

process = psutil.Process()

to_pickle = {
    "type": "FeatureCollection",
    "crs": {"properties": {"name": "EPSG:4326"}, "type": "name"},
    "features": [
        {
            "type": "Feature",
            "id": f"id{idx}",
            "properties": {
                "str_prop": f"str_var{idx}",
                "int_prop": idx,
                "bool_prop": bool(idx % 2),
                "none_prop": None,
                "float_prop": float(idx),
            },
        }
        for idx in range(9000)
    ],
}
b = pickle.dumps(to_pickle)
del to_pickle

print(
    f"Initial memory usage: {round(int(process.memory_info().rss / 1024) / 1024, 3)}MiB"
)
pickle.Unpickler(io.BytesIO(b)).load()
print(
    f"After 1st unpickle memory usage: {round(int(process.memory_info().rss / 1024) / 1024, 3)}MiB"
)
for i in range(1000):
    pickle.Unpickler(io.BytesIO(b)).load()

print(
    f"Final memory usage: {round(int(process.memory_info().rss / 1024) / 1024, 3)}MiB"
)

CPython versions tested on:

3.11

Operating systems tested on:

Linux, macOS

首先在受影响的 CPython 版本上运行提供的 Python 复现程序，并比较重复调用 pickle.Unpickler(...).load() 后的内存。跟踪用于这些嵌套字典的 C Unpickler 路径；当重复反序列化不再使报告的额外内存保持分配状态，并且已为该情况添加回归测试时，即表示完成。