`npy_inflate` fails with "Error inflating stream" on deflate streams larger than ~2 MB

[olokobayusuf]

npzstringreader / npzreader throw std::runtime_error("Error inflating stream") when reading NPZ entries written by numpy.savez_compressed() once the compressed entry exceeds roughly 2 MB (i.e. spans more than one of npy_inflate's 1 MiB input chunks). Small entries read fine. The root cause seems to be the inner loop condition in npy_inflate in src/zip.cpp, which drops unconsumed input.

Repro

Generate a small and a large archive with numpy:

import numpy as np

rng = np.random.default_rng(0)
np.savez_compressed("small.npz", rng.standard_normal(219_040).astype("float32"))  # ~800 KB, reads OK
np.savez_compressed("large.npz", rng.standard_normal(876_160).astype("float32"))  # ~3.2 MB, fails

Read them back with libnpy:

#include <npy/npy.h>
#include <fstream>
#include <iostream>

int main() {
    std::ifstream input("large.npz", std::ios::binary);
    std::string bytes{std::istreambuf_iterator<char>(input), {}};
    npy::npzstringreader reader(std::move(bytes));
    auto tensor = reader.read<npy::tensor<float>>(reader.keys().front());  // throws
    std::cout << tensor.size() << '\n';
}

small.npz reads correctly butlarge.npz throws Error inflating stream.

---

I'm happy to open a PR with a fix. Just say the word.