[Test] Gate extension-requiring tests on actual device capabilities instead of a static arch table.

[duburcqa]

Issue: #

Whether the adstack and data64 extensions are actually usable on the SPIR-V archs (vulkan, metal) is a
property of the initialized device, not of the arch alone: the on-device adstack size interpreter reads through raw
64-bit buffer addresses, so it needs spirv_has_physical_storage_buffer + spirv_has_int64 (the launch gate in
runtime/gfx/adstack_sizer_launch.cpp), and data64 needs spirv_has_int64 + spirv_has_float64. The static
arch-level extension table cannot advertise this, so @test_utils.test(require=...) used to resolve support at
collection time against that table -- which is both too coarse (no device exists yet) and wrong on the SPIR-V archs.

This PR:

• makes is_extension_enabled consult the initialized device's capabilities for adstack / data64 on the SPIR-V
  archs, superseding the static table where the device is the authority;
• moves the test harness's require= resolution from collection time into the wanted_arch fixture: the required
  extensions ride the parametrization and are checked against the actual device after qd.init, skipping with a
  clear reason when the device does not support one.

No behavior change on CPU / CUDA / AMDGPU, where the arch table already matches the device.

Brief Summary

copilot:summary

Walkthrough

copilot:walkthrough

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d382cd7918

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Gate adstack on all SPIR-V sizer capabilities

On a Vulkan device that advertises physical-storage buffers and int64 but lacks int8 or int16, this returns True for qd.extension.adstack, so the updated test fixture will run require=qd.extension.adstack cases instead of skipping them. The actual SPIR-V adstack sizer still hard-errors unless spirv_has_physical_storage_buffer, spirv_has_int64, spirv_has_int8, and spirv_has_int16 are all present (runtime/gfx/adstack_sizer_launch.cpp:332-337), so this check needs to match that runtime gate.

Useful? React with 👍 / 👎.

[github-actions]

Total: 3 file(s) changed, +27 -14 code lines.

[github-actions]

Diff coverage: 73% · 15 lines, 4 missing