Fix SciTS tutorial issues found in an end-to-end Aurora run

[jdwillard19]

• evaluation.md: replace the DDP/HSDP discussion with an overview of the evaluation pipeline (training signals, checkpoints, held-out scoring in the notebook) and a metrics guide; keep limitations short at the end.
• README/environment: the trunk is OLMo-1B, not Qwen3-0.6B; link the notebook; add a recursive clone + submodule step (the env build fails at step 5/7 without the walrus submodule).
• training.md: pass --project (defaults to ModCon).
• data.md: point hackathon users at the pre-staged data and note the ~3 h full-conversion runtime.
• notebook: show the effective max_steps override, hide weight-loading progress bars, install ipywidgets alongside jupyterlab.

Description

Summarize the change and the motivation behind it. Describe the user-visible
behavior, not just the implementation.

Fixes # (issue number)

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that changes existing behavior)
• Documentation update
• Infrastructure, CI, or tooling

Compatibility

PRISM has downstream users running long jobs on HPC systems. Does this PR
change any of the following?

• src.* import paths
• the prism CLI surface
• Hydra config keys
• checkpoint keys
• launcher flags
• None of the above

If any box above is checked, describe the migration path here.

How Has This Been Tested?

List what you ran, and what you intentionally skipped and why. Hardware-specific
results may be reported here when the required platform is unavailable to you.

# e.g.
bash tools/ci/local_quality.sh full
pytest -q -m "not multimodal and not launcher and not integration and not network and not slow and not gpu and not aurora and not perlmutter" tests

• Platform tested on:
• Tests skipped, and why:

New Dependencies, Data, or Assets

Any new dependency, dataset, pretrained model, tokenizer, or other asset — with
its license and provenance. Write "None" if there are none.

Checklist

• I have performed a self-review of my own changes
• The code follows the project's style guidelines (bash tools/ci/local_quality.sh)
• I have commented my code where the intent is not obvious
• I have made corresponding changes to the documentation
• I have added tests that cover this change, or explained why none are needed
• New and existing tests pass locally
• No secrets, credentials, or absolute institutional paths are included
  (see SECURITY.md)

AI/LLM Assistance

See Guidelines for AI/LLM-Assisted Contributions in
CONTRIBUTING.md. All PRs require a human reviewer who is
accountable for the review — a project norm, not yet enforced by branch
protection.

• AI/LLM tools were primarily used to generate code or artifacts in this PR
• I have reviewed all such output line-by-line and understand it

Additional Context

Anything else a reviewer should know.

[pemami4911]

LGTM

Failing unit tests will be addressed in the scits-demo branch.