Add training-fitted normalization to time-series processors

[LIU-kaiyu]

Fixes #491.

Adds optional per-feature z-score normalization to TimeseriesProcessor and TemporalTimeseriesProcessor through normalize_strategy="standard". Normalization remains disabled by default.

The processors incrementally fit the training mean and population standard deviation after resampling and imputation, then apply those fixed statistics to subsequent samples. Each resulting timestep contributes equally; constant features use a scale of 1. Temporal timestamps remain unchanged.

The change also:

• Validates normalization inputs and requires successful fitting before processing.
• Clears previous statistics when refitting, including after a failed refit.
• Preserves fitted statistics through existing serialization and cache fingerprints.
• Adds documentation and a synthetic example showing patient-level splitting before fitting and explicit reuse of training processors for validation/test data.

Validation: 39 focused tests passed in Pixi, including 20 normalization tests covering defaults, imputation, invalid inputs, refitting, serialization, cache fingerprints, and processor transfer. The synthetic example and local contribution checks passed.

A broader run encountered three dataset-backed schema-test failures during synthetic MIMIC loading because of Windows URL/path handling; those are outside this change.