Pythia configs: effective batch 320 + clean one-epoch stop

[Smith42]

The config/pythia-like scaling-study runs target world size 1 and one pass over the Smith42/galaxies train split (N=8,474,566).

Configs (all 12):

• gradient_accumulation_steps 58 -> 54, so the effective batch is 16*20 = 320. One epoch is then floor(N/320) = 26483 steps.
• lr_decay_iters 27000*1.1 -> 26500, so the cosine schedule reaches min_lr right as the single epoch ends (max_iters=30000 stays as an unreached upper cap).

scripts/train.py:

• Graceful end-of-stream handling: the streamed epoch is exhausted near step ~26.2-26.5k, before max_iters. The bare next(tdl) calls in the training loop, estimate_loss and validate now catch StopIteration and stop cleanly (saving a final checkpoint) instead of crashing.
• Data shuffle uses a fixed, rank-independent seed (1337 rather than 1337 + seed_offset). At world size 1 this is a no-op, but it keeps the data order reproducible if these configs are ever run under DDP.

Supersedes #70 (data order is already deterministic at world size 1), #71 (one epoch via batch size + graceful stop rather than max_iters) and #72 (no local sampler needed).

[Smith42]

Solves data ordinance problem by pinning world size to 1

[ksd3]

This is throttling to single-GPU which is fine for the moment