fix: harden generate_voiceover param sanitization

[fancyboi999]

Summary

• make TTS param parsing resilient to invalid LLM outputs (JSON parse and sanitize fallback to {})
• harden type normalization to avoid runtime crashes on malformed values
• keep backward compatibility for legacy numeric range schema encoded as enum: [min, max]

Why

Recent range/enum split improved schema clarity but left edge cases where invalid model output could break the whole voiceover node. This PR ensures field-level degradation instead of pipeline failure.

Validation

• syntax check: python -m py_compile src/open_storyline/nodes/core_nodes/generate_voiceover.py
• regression checks for sanitizer + infer/sanitize paths
• E2E-mock path through process with mocked LLM/provider

Scope

• code only, no docs included in this PR