feat(config): preconfigure GPU accelerator override matrix

[MioYuuIH]

Summary

• add AMD Radeon RX 9600 GRE as a curated accelerator in the default values baseline
• keep the multi-node values example aligned with the default accelerator catalog
• treat 9600 GRE as a curated installer SKU so generated overlays only pin node selectors instead of injecting an ad-hoc stanza
• preconfigure GPU resource image overrides for supported published targets without expanding default visible accelerator options
• make installer overlays emit matching acceleratorOverrides so installer-selected registry/tag values override the preconfigured defaults

Visibility model

• custom.accelerators is the curated accelerator catalog.
• metadata.<resource>.acceleratorOverrides predefines image routing for supported targets.
• metadata.<resource>.acceleratorKeys remains the deployment/admin-controlled visible set.
• Defaults stay conservative: GPU resources still expose only strix-halo unless an installer or environment overlay adds more keys.

Override matrix

• phx -> latest-gfx110x
• strix -> latest-gfx1150
• strix-halo -> latest-gfx1151
• 9070xt/r9700/9600gre -> latest-gfx120x

Installer behavior

• Installer-generated overlays now emit acceleratorOverrides for every detected/enabled accelerator, including homogeneous single-target installs.
• This prevents default latest-* overrides from replacing installer-specific images such as develop-gfx1151, custom registries, or offline bundle tags.

Compatibility notes

• Deployments that override GPU resource images with custom tags or registries must also override matching acceleratorOverrides entries.
• Installer-managed installs and upgrades now generate those matching entries automatically.
• Manual Helm overlays can use the preconfigured matrix as a template, but should keep acceleratorKeys limited to accelerators actually exposed to users.

Validation

• pytest tests/installer/test_gpu.py tests/installer/test_overlay.py tests/installer/test_values_gpu_overrides.py
• YAML parse check for runtime/values.yaml and runtime/values-multi-nodes.yaml.example
• git diff --check

Notes

• This keeps generated values.local.yaml unchanged in the repo.
• The PR intentionally does not make the frontend show every curated accelerator by default; admins opt in by adding keys to acceleratorKeys.

[MioYuuIH]

Closing this draft while we reconsider how much GPU accelerator coverage should be exposed by default versus filtered by installer/environment overlays.