print experiment hyperparams in final summary

[mvanhorn]

The final summary block prints output metrics (val_bpb, peak_vram_mb, etc.) but only one input parameter (depth). Agents running overnight sessions re-read train.py after every experiment just to recall what learning rates or batch size they used.

This adds the key hyperparameters to the existing summary format - seed, batch sizes, learning rates, weight decay, and window pattern. Same key: value layout, no new files or imports.

Complementary to #331 (which covers structured output metrics via results.json - this covers input config via stdout).

Before:

num_params_M:     50.3
depth:            8

After:

num_params_M:     50.3
depth:            8
seed:             42
total_batch:      524288
device_batch:     128
matrix_lr:        0.04
embedding_lr:     0.6
weight_decay:     0.2
window_pattern:   SSSL

This contribution was developed with AI assistance (Claude Code).