feat(wgpu): add guarded host-visible primary memory

[shinaoka]

Summary

• add opt-in PrimaryMemoryMode::HostVisible with whole-resource exclusive primary allocations
• add guarded read/write mapping, stable allocation identity, lease-safe clones/guards, and atomic GPU/host exclusion through submission
• add repeatable sync/async graphics-API device initialization for independent Metal contexts

Verification

• Metal host-visible integration: 2 passed
• filtered package suite: 375 passed, 14 ignored, 4 known baseline WGSL-U8 tests filtered
• fmt, clippy -D warnings, std+msl check, rustdoc -D warnings
• independent task re-review: no Critical/Important findings

Notes

The four unfiltered WGSL-U8 failures predate this branch and their sources are unchanged. Crate publication is intentionally deferred; tenferro/CubeK may pin the reviewed Git commit for now.

Supports tensor4all/tenferro-rs#1394.