docs(experiments): add PyTorch video poker example with a full W&B run lifecycle

[jstack-wandb]

What

Adds examples/pytorch/pytorch-video-poker-bot/, a PyTorch example that trains a small network to play 9/6 Jacks or Better video poker and tracks training and evaluation in W&B. No existing example changes.

Why

Most examples in this repo are several years old, and a newer example was requested that shows current W&B usage end to end. The poker task is a stand-in: the README is written for readers learning W&B, not poker.

How

• Three scripts run in order: generate_dataset.py builds the training data, train.py trains and logs, and evaluate.py plays hands with the trained model and logs the score.
• wandb.init() records the config, puts each run in a group (train or eval), and takes an optional --entity for logging to a team.
• run.log() logs metrics every epoch in training and running totals every 100 hands in evaluation.
• run.log_artifact() uploads the best checkpoint as a versioned model artifact. evaluate.py downloads it with run.use_artifact(), which links the evaluation run to the training run in the artifact's lineage.
• Runs finish when the with wandb.init(...) block exits; the README explains that run.finish() is needed without it.
• File paths and W&B names are not hardcoded. The dataset path, checkpoint path, artifact name, project, and run name are required flags.
• Seeds are random unless --seed is passed. The seed used is saved in the run config, so any run can be replayed.
• The best checkpoint is the epoch with the lowest val_mean_regret: the average expected profit per credit bet that the network's holds give up compared with the optimal hold.
• The training data is not committed. generate_dataset.py writes 1M hands (about 44 MB) in about two minutes. Each hand is labeled with the exact expected profit of all 32 hold choices, which ev.py computes by counting every possible draw from precomputed tables instead of simulating. .gitignore covers generated data, checkpoints, downloaded artifacts, wandb/, and .venv/.

Testing

• Ran the README steps against a local W&B server and against QA (lucas-defino-weights-biases/pytorch-video-poker): 5 training runs with random seeds, then 5 evaluations of the lowest-regret model at 250,000 hands each.
• QA training runs reached a best val_mean_regret of 0.0005 to 0.0014 and 98.8% to 99.2% validation return.
• The five QA evaluations returned 99.14% combined over 1.25M hands, against about 99.5% for optimal play. Individual runs ranged from 98.34% to 99.71%. Most of that spread comes from royal flushes, which pay 4,000 credits and occur about once per 40,000 hands.
• ev.py matched brute-force enumeration of every possible draw for 3 random hands and all 32 holds, with zero difference.
• Passing a run's logged seed back with --seed reproduced its training metrics exactly.
• Not tested: Python versions other than 3.9, and GPUs. The code runs on CPU only.

Anything Else

• New dependencies: requirements.txt lists numpy, torch, and wandb with no version pins, like 15 of the other 21 requirements*.txt files under examples/. Tested with numpy 2.0.2, torch 2.8.0, and wandb 0.26.1.
• evaluate.py downloads the model with run.use_artifact(), so it does not run with WANDB_MODE=offline.
• Should requirements.txt pin the tested versions instead of matching the unpinned majority?