Add structured and tabular data redaction

[jzonthemtn]

Summary

Phileas is built around free text. It does not have a first-class path for structured and semi-structured data (CSV, database tables, JSON) where redaction should be column- or field-aware. Microsoft Presidio offers this via presidio-structured. This issue tracks adding structured/tabular support to Phileas.

Surfaced while writing the Phileas vs Presidio comparison, where we concede structured/tabular data to Presidio.

Why it matters

A large share of PII lives in tables and JSON, not prose: exports, warehouse columns, event payloads, API logs. Field-aware redaction (detect that a column is SSN, apply a strategy to the whole column consistently) is materially different from running free-text detection over a serialized blob, and it pairs naturally with the AI training-data and analytics use cases where consistent pseudonymization matters. Without it, structured-data teams have a reason to pick Presidio.

Acceptance Criteria

• Phileas can accept structured input (at minimum tabular/CSV and JSON) and apply redaction per field or column rather than only as free text
• Detection can be column-aware: infer the PII type of a column from a sample and/or accept an explicit field-to-entity mapping
• Field-level strategies reuse the existing policy model and strategies (redact, mask, FPE, consistent pseudonymization), so a value pseudonymized in one column maps consistently where intended
• Consistency across rows is supported (the same input value maps to the same token within a configured scope), which is the key requirement for analytics and training data
• The capability is exposed in at least the primary (Java) runtime with a documented path for Python
• User documentation under docs/docs/ covers structured/tabular redaction with a runnable example
• The Phileas vs Presidio comparison and product page are updated once shipped