Add per-CPU memory cap support for SLURM clusters

[mpvanderschelling]

Summary

Follow-up to #352 / #353 (per-CPU memory accounting). Adds a cluster-level
per-CPU memory cap so an entire pipeline is submittable on sites whose
job_submit filter rejects --mem and caps per-CPU memory (e.g. TU Delft's
DelftBlue, 3968 MB/core).

SlurmCluster.mem_per_cpu becomes str | None carrying the cap (e.g.
"3968M"). When set, every generated script (orchestrator + steps) emits
--mem-per-cpu derived from each step's declared per-node mem:

• per-CPU value = ceil(mem / cpus_per_task)
• cpus_per_task is bumped up to ceil(mem / cap) when the node's core:memory
  ratio can't supply the declared memory on the requested cores (the only way to
  buy more memory on such a cluster is more cores). It is only ever increased,
  never reduced.

An explicit per-resource SlurmResources.mem_per_cpu (from #352) still renders
verbatim and wins. Clusters that accept per-node --mem leave the cap None —
behaviour unchanged (e.g. Brown's Oscar).

Why

With only the per-resource field from #352, every experiment/step would need a
manual DelftBlue-specific override, and the library-default orchestrator
resources (--mem=1G) were themselves unsubmittable there. This keeps
total-memory declarations cluster-agnostic: each cluster maps them to its own
core:memory ratio.

Tests

Added 8 cases to TestPerCpuMemory covering the cap fit / bump / divide paths,
explicit-override precedence, orchestrator rendering, and the default-None
(Oscar) path. Full test_slurm.py + test_resources.py pass (57).

🤖 Generated with Claude Code