orchestrator: use block-sized memfile fetches behind a feature flag

[matthewlouisbrockman]

Cold resume on the memfile path could fetch 4 MiB even when the requested block size was smaller, such as 2 MiB. This change lets memfile fetches align with the actual block size, reducing over-fetching. Charts to follow but draft for sharing

nvm, results in less memory loaded but slower cause of gcs reads

[cursor]

PR Summary

Low Risk
Low risk because behavior changes are gated behind the new memfile-block-sized-fetch feature flag and only affect memfile reads; main risk is subtle cache/fetch boundary mismatches leading to unexpected read errors or performance regressions when enabled.

Overview
Adds an opt-in feature flag to make the non-streaming chunker fetch memfile data in blockSize increments instead of the fixed storage.MemoryChunkSize, by threading the storage object type into NewChunker and introducing a configurable fetchSize used throughout FullFetchChunker reads/writes.

^{Written by Cursor Bugbot for commit 164f747. This will update automatically on new commits. Configure here.}