FusedCodecPipeline: forked worker hangs on a multi-chunk read; reset_resources_after_fork misses codec_pipeline._pool #4478

Versions: zarr 3.4.0, numcodecs 0.17.0, numpy 2.4.6, Python 3.12.14, Linux.

Description. With codec_pipeline.path = FusedCodecPipeline and max_workers > 1 (the
default), multi-chunk I/O in a parent creates the module-level zarr.core.codec_pipeline._pool.
reset_resources_after_fork (zarr/core/sync.py) resets loop, iothread and _executor in
a forked child, but not _pool / _pool_size, so the child reuses an executor whose threads do
not exist. submit consumes an idle permit the parent left rather than starting a thread; a read
of no more chunks than the inherited permits never returns from read_sync's pool.map (a
larger read happens to start a thread and completes, so the hang depends on read size).

Reproducer (fork start method, the Linux default before Python 3.14):

import multiprocessing as mp, os, sys, tempfile
from concurrent.futures import ProcessPoolExecutor, TimeoutError
import numpy as np, zarr
import zarr.core.codec_pipeline as cp

cfg = {"codec_pipeline.path": "zarr.core.codec_pipeline.FusedCodecPipeline"}
if len(sys.argv) > 1:
    cfg["codec_pipeline.max_workers"] = int(sys.argv[1])
zarr.config.set(cfg)
path = os.path.join(tempfile.mkdtemp(), "a.zarr")
a = zarr.create_array(path, shape=(40, 40), chunks=(10, 10), dtype="i2")
a[:] = np.arange(1600, dtype="i2").reshape(40, 40)  # 16 chunks: parent uses cp._pool
a[:]
print("zarr", zarr.__version__, "| parent pool idle permits:",
      cp._pool._idle_semaphore._value if cp._pool else None)

def child(_):
    return int(zarr.open_array(path, mode="r")[:10, :20].sum())  # 2 chunks

pool = ProcessPoolExecutor(2, mp_context=mp.get_context("fork"))
try:
    print("children:", list(pool.map(child, range(2), timeout=30)), flush=True)
except TimeoutError:
    print("HANG: forked workers did not return in 30 s", flush=True)
    for proc in pool._processes.values():
        proc.kill()
    os._exit(1)

Output:

$ python repro.py
zarr 3.4.0 | parent pool idle permits: 10
HANG: forked workers did not return in 30 s
$ python repro.py 1   # codec_pipeline.max_workers = 1
zarr 3.4.0 | parent pool idle permits: None
children: [37900, 37900]

Hangs in 3 of 3 runs with the default pool; completes in 3 of 3 with max_workers = 1. We first hit
this on Zarr v2-format arrays with a numcodecs Blosc compressor, in a fork-pool build reading
multi-chunk bands.

Suggested fix: reset _pool = None and _pool_size = 0 in the child (in
reset_resources_after_fork, or an os.register_at_fork in codec_pipeline.py).
Workaround: codec_pipeline.max_workers = 1, which never creates the pool.