feat: nightly hardening - strict configuration validation

[mouse-value-add]

Problem

PromptLens accepted invalid runtime configuration values (e.g., out-of-range temperatures, non-positive token limits, unsupported output formats), which could fail late at runtime and make evaluations brittle.

Approach

• Added strict Pydantic validators across config models to fail fast on invalid values.
• Enforced sane boundaries for temperatures, token/timeouts, and retry settings.
• Added output format normalization + deduplication (markdown alias -> md) with explicit allowlist checks.
• Enforced non-empty model configuration at run-level.
• Added focused tests for validation behavior and normalization.
• Documented the hardening capability in README features.

Verification

• Ran: python3 -m pytest -q
• Result: 5 tests passed.

Risks

• Some previously accepted but invalid config values now fail early with validation errors.
• Users relying on permissive/typoed output formats will need to update configs.

Rollback Plan

• Revert this PR commit to restore previous permissive config behavior.
• If needed, selectively relax individual validators while keeping test coverage.