Strict/warn/lenient failure semantics; remove blanket warning suppression

[marwinsteiner]

Context

calibration.py sets a module-level warnings.filterwarnings("ignore") and compute_ivs_vectorized catches bare Exception into NaN. Failed slices are skipped. Convenient, and exactly how data problems become invisible.

Proposal

• Explicit failure semantics: mode="strict" | "warn" | "lenient" on the ingestion/fit entry points. Strict: bad input raises. Warn (default): current behaviour plus recording. Lenient: filtered and recorded. Nothing disappears silently -- rejected quotes and failed inversions become counts/locations on the fit report.
• Remove the blanket warnings filter; catch only the specific py_vollib exceptions.

Acceptance criteria

• Strict mode raises on the first bad quote/slice with location
• Rejection accounting flows into the fit report in all modes
• No module-level warning suppression remains