Clarify Speedrun starter discovery for new contributors

[esaran1]

Speedrun onboarding_setup.sh referenced a missing hackable starter template and
a broken submitting-speedrun doc link. Fall back to train_125m_fineweb_edu_gpu.py,
document --dry_run true, and note that the script requires a GPU (not Mac).

[dlwh]

Hi @esaran1 we're kinda moving away from speedrun for now and toward the agent-moe experiemnt gates for most things. What are you most interested in speedrun?

[dlwh]

(I recognize we should clarify this!)

[esaran1]

Thanks for the context.

I hit the broken speedrun onboarding path while trying to get started as a new
contributor (missing hackable_transformer_starter_template.py). This PR is
mostly meant to unblock that discovery issue, not to push a specific speedrun
leaderboard submission yet.

My underlying interest is compute-efficient LM training: reproduce a clean
baseline, then test controlled training-recipe changes (LR schedule, AdamW
settings) with fair BPB/FLOP comparisons before scaling. I started with
speedrun because that was the public onboarding direction, but I'm happy to
align with the current canonical path if that's Grug/MoE gates now.

Could you point me to the recommended entry point for a new contributor today?
(e.g. experiments/grug/base or the moe gate workflow). I'm on Mac for dev and have HPC access.