Improve performance: memory footprint and speed

[j-atkins]

This PR makes changes to fieldset ingestion to improve performance, especially so in RAM-limited workspaces.

• Adds more FetchSpec detail to different instruments (principally for the depth dimension), to minimise redundant data retrieval.
• Move to using Parcels to_chunk_cached_arrays() instead of to_windowed_arrays(), which is more suitable and performant for VirtualShip simulations.
  • Also sets the max_cache_bytes to 300_000_000 (rather than the Parcels default 600_000_000).
  • This appears to strike a balance of speed vs. memory footprint, which I have found to be especially important when using VirtualShip on RAM-limited virtual machines, e.g. the VirtualShip GitHub Codespaces implementation.
• Improved performance with to_chunk_cached_arrays() means the tmp ds writing step (_via_tmp_ds) can be removed in VirtuaShip. Therefore, wrapping instrument simulations in a context manager for proper tmp dir handling is also removed.

---

N.B.

There is a temporary fix in argo_floats.py to handle an indexing bug in Parcels (i.e. #2906). I will keep this PR as a draft until fixed upstream.

---

• Update tests