Optimize CSS visibility engine with lazy parsing and cache-friendly evaluation

[arrufat]

This PR brings significant performance improvements to the new CSS engine by shifting to a lazy architectural model and optimizing hot-path lookups for cache locality.

• Lazy Architectural Model:
  • Deferred the allocation of heavy JS-backed CSSStyleRule and CSSStyleDeclaration objects. Stylesheets are now only parsed into the full DOM object tree on-demand (e.g., if queried by Javascript).
  • Implemented a fast-path raw text parser for the StyleManager, allowing it to build its internal visibility rules without triggering thousands of JS object allocations during page load.
• Cache-Friendly Data Structures (SoA):
  • Refactored internal rule storage to use Structure-of-Arrays (SoA) via std.MultiArrayList.
  • This allows the evaluation loop to iterate over lightweight priority integers in a dense cache line, fast-skipping irrelevant rules without loading heavier property or selector data.
• Simplified Priority Logic:
  • Combined CSS specificity and document order into a single packed u64 priority.
  • This replaced complex tie-breaking logic with a single, native > integer comparison, simplifying the code while increasing speed.
• Early Rejection:
  • Implemented early exits in evaluation loops for elements already determined to be hidden by higher-priority rules or inline styles.

Impact: Closed the performance regression on large page fetches by around 1 second (from 7.5 sto 6.5 s) while the main branch does it in about 6 s (but doesn't compute the styles correctly)