Restore z-score normalization for pi0_fast_libero configs

[0xadvait]

Problem

Evaluating the official pi0_fast_libero checkpoint on current main gives ~0% success on LIBERO (the arm spins in place). #849 has six independent reports, and one user bisected the last working commit to e458066.

Root cause

At e458066, LeRobotLiberoDataConfig built its DataConfig without setting use_quantile_norm, so the pi0-FAST LIBERO configs trained and served with z-score normalization. Current main's DataConfigFactory.create_base_config force-sets use_quantile_norm = model_type != PI0, which silently flipped pi0_fast_libero and pi0_fast_libero_low_mem_finetune to quantile normalization. The released checkpoint was trained with z-score stats, so the state input is mis-normalized going in and the action chunk is mis-unnormalized coming out. The scale mismatch on actions is what produces the spinning-arm behavior. DROID FAST configs are unaffected; they used quantile normalization both before and after that change.

Evidence

Deterministic probe: the official checkpoint, a fixed synthetic LIBERO observation, greedy FAST decoding, run at e458066 and at current main. Preprocessed image hashes and transform sequences are identical at both commits, but:

• post-normalization state vector differs by up to 10.36
• FAST token ids diverge at 5 positions
• action chunks differ with max abs diff 2.64

With this PR applied to main, the same probe reproduces e458066's post-normalization state, token ids, and both action chunks exactly (max abs diff 0 on all four checks, two infer calls each).

Change

• DataConfigFactory gains use_quantile_norm_override: bool | None (default None keeps the model-type default for every existing config).
• The two pi0_fast_libero configs pin the override to False with a comment explaining the checkpoint compatibility constraint.

Flag matrix after the change: pi0_fast_libero False, pi0_fast_libero_low_mem_finetune False, pi0_fast_droid True, pi0_libero False, pi05_libero True.

ruff check and ruff format --check pass. data_loader_test.py: 4 passed, plus test_with_real_dataset which fails identically on clean main (pre-existing dataset-fetch issue, unrelated).

Caveat

Anyone who finetuned pi0-FAST on LIBERO with current main trained under quantile normalization; for those checkpoints the override makes the choice explicit and per-config instead of a silent default flip.

Fixes #849