Split from #747. Issue for discussing an mmap primitive being used for WebGPU buffer mappings.

Presentation given 2026-06-24 discussing Chromium's prototyping with V8+Blink.

Discussion points listed at the end of that presentation:

• memcpy is pretty fast.
• Are there applications that use enough bandwidth that this is worth optimizing? Enough relative to CPU/GPU time?
• Should we pursue persistent mapping?
• Is it worthwhile to compromise with mprotect?