Add client-side W&B logging example for Tinker RL loops

[micahtyong]

Summary

Adds scripts/wandb_rl_example.py: a minimal multi-step Tinker RL loop against a Lilo deployment that logs per-step metrics to Weights & Biases, plus a README pointer. Answers the "how do you use W&B with Tinker code?" question — W&B is entirely client-side; the Lilo server never sees it.

Per step: save_weights_and_get_sampling_client() → sample a group → toy binary reward → forward_backward(datums, "importance_sampling") + optim_step(AdamParams) → wandb.log({reward/*, response_len/mean, time/*, train/<metrics Lilo returns>, optim/*}, step=step).

Two details that matter against the deployed backend: create_lora_training_client(..., train_unembed=False) (matches the deployment's LoRA targets), and target_tokens = prompt[1:] + tokens so it lines up exactly with model_input = prompt + tokens[:-1].

Validated with a 3-step run against the lilo deployment (reward 1.0 each step, ~14 s/step after warm-up): https://wandb.ai/modal-labs/lilo-examples/runs/zg7rnr16. ruff check/ruff format --check pass.

tinker-cookbook users don't need this: set wandb_project/wandb_name on the cookbook Config (that's what the LongRLVR runs in #30 do).

Link to Devin session: https://modal.devinenterprise.com/sessions/8d6804dd317c41329d77eb2c427dda6f
Open in Devin Desktop: https://modal.devinenterprise.com/desktop/session/8d6804dd317c41329d77eb2c427dda6f?variant=devin
Requested by: @micahtyong

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[kailash109]

lgtm!