fix(experimental): canonicalize typed and dynamic graph buffer imports

[ibgreen-openai]

Goals

Preserve one canonical command-graph handle when the same caller-owned GPU allocation is imported through raw buffers, typed GPU table data, or resizable DynamicBuffer wrappers.

Changes

• Reuse an existing raw-buffer handle for later GPUData imports instead of creating independent logical resources for the same allocation.
• Track DynamicBuffer wrapper identity across backing-buffer replacement while retaining the original raw-buffer canonical handle.
• Preserve current master's native-GPU-buffer physical alias detection, safe read-only sharing, and writable-alias rejection.
• Add focused real-WebGPU regressions for import ordering, wrapper identity, resized backings, borrowed ownership, and repeated graph encoding.

Verification

• Full cross-package yarn build passes.
• Focused actual Chromium/WebGPU graph, sorting, Arrow, similarity, clustering, IVF, and benchmark coverage: 88 tests passed.
• Focused Node regressions across graph, tables, Arrow, luVS, and docs: 99 tests passed.

Landing order

This is the first, two-file prerequisite. Merge it directly into master before reviewing or merging the fixed-size-list table PR.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 33e233fa1f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".