Fix SimpleDataSet overriding the configured RNG seed

[sopenlaz0]

Summary

Trainer seeds Python's process RNG from Global.seed, but SimpleDataSet
immediately calls random.seed(None) while sampling and shuffling whenever the
dataloader does not pass an explicit dataset seed. That replaces the configured
seed with system entropy, so otherwise identical runs can start with different
sample orders.

This change keeps the existing process RNG state when seed is None, while
preserving the previous explicit-seed behavior.

The fix is intentionally scoped to SimpleDataSet. It does not reinterpret
the existing seed argument globally, because other datasets use None and
epoch for different scheduling semantics.

Verification

• Added a behavioral test using the real SimpleDataSet path:
  • same process seed -> identical sample order
  • different process seed -> different sample order
  • explicit dataset seed overrides process state
• pytest -q tests/test_dataloader_seed.py: 1 passed
• Structured autoreview: clean, no accepted/actionable findings

Base commit: 0d522801ec6dc1df852c6b6d4ed6a08f5127ed97.