e2e: the ReproNim/containers clone should be an optional cache, not a prerequisite

## What

The e2e suite's two SimBIDS app configs (`mechababs/testing/e2e/pipelines/SimBIDS-0.0.3+{anchor,chain}.yaml`) carry `container.source: ../containers`, a path that resolves against the study root, so the suite's studies and a hand-seeded ReproNim/containers clone must share a parent (`MECHABABS_E2E_WORKDIR`).
`conftest.simbids_sif` skips the whole suite if the SIF is not already there, and the seed is a two-command recipe the skip message prints.

The shipped example config (`examples/bids-app-configs/SimBIDS-0.0.3.yaml`) names the same dataset by URL, `https://github.com/ReproNim/containers.git`, which is what a production config carries and what `scaffold.container_source` passes through verbatim.
So the e2e exercises a different `container.source` shape than the one users are told to write.

## Why the path is there

babs installs `container.source` into every derivative it inits.
With a URL that is a fresh clone of ReproNim/containers plus a `datalad get` of the SIF per cell, per test run; the sibling clone makes it a local clone from an already-populated dataset.
It is a cache, and it is worth keeping as one.

## Fix

- Default the e2e configs to the URL, matching the example.
- Add an env-var override (`MECHABABS_E2E_CONTAINERS=`) that the fixtures substitute into the app configs when set, so a local clone anywhere serves as the cache. The path need not be a sibling of the studies; the `../containers` assumption goes.
- `simbids_sif` stops being a skip gate: with no override, babs does the pull; with one, the fixture checks the SIF is present there.
- `run_in_podman.sh` forwards the variable and bind-mounts the path when set, the way it already handles `BABS_SPEC` and the workdir.

Cost without the override: one clone of ReproNim/containers per cell in CI (~4 cells across the three scenarios). Measure it once; if it dominates the ~3 min podman run, the podman script can seed the cache itself instead of the user.

## Context

Austin, 2026-08-22, at the chunk-3 review: "let babs do the pull itself; add an env-var override for quick tests; kill the `../containers` sibling assumption."
The examples half of that is done (#133 moved them to the ReproNim URL); this is the e2e half.

Data point, 2026-09-03 (con/mechababs#138, the CI workflow): the seed step costs 0 s on an `actions/cache` hit (4 s restore, ~330 MB) and 28 s of host-side git-annex + datalad install that exists only for seeding; the image pull is 38 s regardless.
Austin, same day, asked about a URL-for-CI + env-var override and then dropped it ("nevermind"); the CI-side argument against was per-cell clone + SIF fetch inside the container vs a cached one-time seed, and that the seed is verbatim `test-cluster`'s documented host prep.
The 2026-08-22 directive above stands as written.

Start with the two SimBIDS e2e configs, conftest.simbids_sif, and run_in_podman.sh; compare them with examples/bids-app-configs/SimBIDS-0.0.3.yaml and the existing BABS_SPEC/workdir handling. Trace how container.source is substituted and how the SIF check currently gates tests. Done means the URL is the default, MECHABABS_E2E_CONTAINERS works from any path, and podman forwards and mounts it when set without skipping otherwise.

Written by the indexing model from the issue text.