Config composition leaks: cluster/runtime settings land on the pipeline axis

## Problem

A run composes three axes — dataset × pipeline × cluster — but today cluster- and container-runtime settings leak onto the **pipeline** axis, where they don't belong: SLURM partition, templateflow/FreeSurfer-license host binds, container flags (`--containall`, `--userns`, `$JOB_TMP:/tmp`). A pipeline YAML that hardcodes a site's paths/partition can't move to another cluster without editing it.

## Direction (not settled)

Treat the invocation as a **need↔provider join**, not a flat merge: the pipeline declares abstract *needs* (a mountpoint, a partition, writable `/tmp`) with no host paths; the cluster declares *providers* (the host path, the partition name); the composer joins them, and code branches on capability, never on cluster identity.

- Some of this is already babs-native (templateflow, fs-license) — we're double-handling.
- A generic host-file need/provider mechanism might be worth pushing to babs (an `EXTRA_FILES`-style config), complementing `PennLINC/babs#385` (`common_paths`) for files outside the sparse-checkout.
- `merge_config` composing cluster × pipeline (instead of overwriting) is the mechababs-side lever.

Fuzzy for now — the design is in progress; this issue is the tracker.

This is a design tracker with no files or tests named. Start by reading the need↔provider direction and locating the merge_config composition mentioned in the issue. Done means agreeing on and implementing a settled separation of pipeline needs from cluster providers, including the listed runtime settings, with the design no longer marked fuzzy.

Written by the indexing model from the issue text.