Add OpenCode and Grok Build adapters

[starkdmi]

What changed

This adds the new OpenCode and Grok Build adapters and folds in the follow-up fixes that came out of review and real-data validation.

Why it changed

The goal is to ingest real OpenCode and Grok Build usage, recover per-model attribution where the raw data is incomplete, and estimate API-equivalent cost more fairly for provider-qualified and proxy-wrapped model names instead of leaving them at $0.

Impact

• adds OpenCode and Grok Build source scanning
• improves OpenCode model recovery from message records and preserves provider-qualified model identity
• splits ambiguous multi-model OpenCode sessions when message-level usage is available, with residual aggregate fallback when reconstruction is partial
• broadens pricing normalization for provider-prefixed and proxy-wrapped model IDs
• makes Grok unified-log handling incremental per session for cache invalidation
• hardens Grok JSONL parsing so malformed rows are skipped instead of aborting the whole scan

Root cause

The initial adapter implementation handled the core formats, but real local data exposed a few gaps:

• OpenCode session-level model fields are often missing even when message-level model data exists
• proxy/local-provider model names need normalization to map onto known pricing tables
• Grok JSONL files can contain malformed trailing or partial rows that should not block usage collection

Validation

• cargo test -p statsai-adapters -p statsai-pricing -p statsai
• real-data preview scans for OpenCode and Grok Build during development