Fix token sampling underflow for short token datasets

[TastyHeadphones]

Summary

• add a dataset length guard after mmap in train_large.m
• prevent size_t underflow in max_pos = n_tokens - SEQ - 1
• return early with a clear error when the token file is too short for one training window

Why

max_pos is computed from unsigned size_t. If n_tokens <= SEQ + 1, the subtraction wraps to a huge value, which can produce invalid random offsets and out-of-bounds reads.

Validation

• built successfully with make train_large in training/

[maderix]

Thanks for catching this — clean fix for a real underflow bug. Merged!