Reading and writing separately from compute in the chunk executor #4860

In TimeSeriesChunkExecutor, each worker reads a chunk, computes on it and writes it. So n_jobs sets three things at once: how many reads are in flight, how many cores compute and how much RAM is used. Storage backends want very different numbers of reads in flight, and that number has nothing to do with the compute.

#4833 and #4347 change the unit of work and where it runs; this is about how reads and writes are issued, and would work with either.

Backends want different numbers of reads in flight

A spinning disk serves one sequential reader best, since concurrent reads make it seek. NVMe wants the opposite: cold-cache reads of 384 int16 channels from a WD SN850X, with N threads each copying the next 1 s chunk out of a memory map as the executor does, ran at 5.6–5.7 GB/s on one thread and 6.8–6.9 GB/s on 8 or 16.

Compressed formats differ too. Reading 1 s chunks with N threads sharing one open file (40 s on NVMe, chunks of 3,750 × 384):

h5py holds a global lock, so threads don't help even with decompression; zarr's codecs release the GIL and scale 5.5 times on 8 threads.

Writes have the same coupling

Each worker writes its chunk at its own offset, so on a spinning disk the writes interleave too. With zarr v3 shards, several workers write parts of one shard; #4831 and #4833 address this by giving each job a run of chunks that covers a shard.

Proposal

The executor would run three stages:

1. A reader issues reads the way the source prefers, into a window bounded by a memory budget. A source could declare a read size, an alignment (such as zarr chunk boundaries), how many reads to keep in flight and whether to read in time order. Sources that declare nothing are read as today.
2. Compute takes chunks from the window and runs the chain. Adjacent chunks share margins, which are already in the window, so nothing is read twice.
3. A writer writes results in time order, and could gather a whole zarr shard before writing it, whatever the chunk size.

No preprocessing step would change. With threads inside each step (#4857), one compute worker can use every core, so read concurrency, compute threads and RAM become separate settings.

Questions

1. Should the executor own reading and writing like this, or should that stay with the extractors?
2. Should read preferences live on the recording segment, since the extractor knows its store, or in job kwargs, since the user knows the disk?
3. Would an opt-in reader for binary recordings be a useful first step?