Add full provider input schemas for all video models

[SecurityQQ]

Context

PR #194 added model-rules.ts with per-model Zod schemas to normalize duration values before sending to providers. Currently ModelDurationRules only handles duration — but there are many other model-specific input constraints in the if/else chains in fal.ts (lines 520-720).

What needs to happen

Expand model-rules.ts into full provider input schemas that validate and normalize all model-specific fields:

• aspect_ratio — some models use video_size instead, some restrict to specific ratios
• resolution — Grok Imagine defaults "720p"
• generate_audio — Kling v3/v2.6 and LTX-2 default to true
• delete_video — Sora 2 needs false to keep URLs accessible
• video_size — LTX-2 uses this instead of aspect_ratio
• num_frames — LTX-2 converts duration * fps to frame count
• end_image_url — supported by Kling v3/v2.6 and LTX-2

Once complete, rename ModelDurationRules → ModelInputRules and consolidate the remaining if/else logic in fal.ts into the Zod schemas.

Related

• Gateway companion issue: vargHQ/gateway#88
• Current implementation: src/ai-sdk/providers/model-rules.ts