Hosted platform tool: optional augmentation, PII scrub, token metering

This release makes SEED runnable as a hosted, governed platform tool (e.g. an MCP tool on AgentCore: Alpine/arm64, Python 3.14, budgeted, PII-sensitive).

Breaking changes

• Image augmentation is now an optional extra. If you use --augment / augment=True, install pip install "seed-data[augment]" (included in [all] and [dev]). Without it, augmentation stops with an install hint before any model call. This keeps opencv out of the base install, so it builds on Alpine/musl.
• A saved InferredSchema JSON with duplicate entity names (which earlier versions could write when planning from several inputs) now fails to load. Merge or rename the duplicates.

New

• seed_data.scrub.scrub_schema(schema, detector): removes PII that a schema planned from real files can echo back (enum values, defaults, examples, patterns, reference samples, free text). It works with any detector (e.g. Bedrock Guardrails, Comprehend), fails closed, and its report never contains values.
• seed_data.token_meter(): measures token usage across every model call. StructuredResult.token_usage is now filled.
• Generator.plan_scenarios() + generate_batch(scenarios=...): plan once, then generate in budget-checked waves.

Fixes

• Planning from several inputs (e.g. a CSV plus a description) no longer produces duplicate entities that overwrite each other's tables. Same-name entities are merged, columns merge case-insensitively, and foreign keys keep pointing at the right table.
• The deprecated strands_tools.calculator is replaced by a sandboxed arithmetic calculator (no shell access, bounded inputs).
• Version floors strands-agents>=1.45 and strands-agents-tools>=0.8 prevent installs from silently falling back to an empty placeholder release.
• Dependency updates: mcp 1.30.0, multidict 6.9.1.