Split neural_lam/utils.py into a utils/ package

[Sir-Sloth-The-Lazy]

neural_lam/utils.py has grown past 800 lines and mixes unrelated concerns graph loading/zero-indexing, MLP construction, plotting/LaTeX detection, training-logger setup, tensor math, and time conversion. New work (e.g. the latent encoder/decoder infra) adds more graph helpers and has had to introduce a TYPE_CHECKING, guarded BaseDatastore import to dodge a circular import.

Proposal

Turn utils.py into a utils/ package, mirroring the models/ layout, with one module per concern:

• buffer_list.py : BufferList
• graph.py : load_graph, zero_index_*
• networks.py : make_mlp
• plot.py : has_working_latex, fractional_plot_bundle
• logging.py : setup_training_logger, init_training_logger_metrics, log_on_rank_zero
• tensor.py : inverse_softplus, inverse_sigmoid
• time.py : get_integer_time

utils/__init__.py re-exports the full public API so existing imports (from neural_lam import utils, from neural_lam.utils import load_graph, …) keep working unchanged. This should also let graph-related helpers sit next to BaseDatastoreaware code and remove the TYPE_CHECKING workaround.

Notes

• Raised from review feedback on the latent encoder/decoder PR feat: add latent encoder/decoder infrastructure for Graph-EFM port #648 ; tracked separately so the refactor doesn't bloat that PR.
• Pure code movement, no behavioural changes, happy to take this on, or for someone else in the community to pick it up.