fix(dashboard): allow re-adding a disabled ACP runner key

[fengyue-xve]

Summary

Fixes #1815. The ACP runner create/edit form rejected any existing runner key with runnerKeyExists, so once a custom runner had been disabled there was no way to add it back under the same name — and disabled custom runners are exactly the state a user is most likely to re-add. Builtin runners cannot be deleted, so for them the form was a dead end too.

The save guard now only treats enabled or builtin keys as collisions:

• disabled custom key → the save overwrites the stale entry; the create form defaults to enabled, so the runner comes back enabled
• enabled custom key → still rejected with the same toast as before
• builtin key (even a disabled one) → still rejected, so builtins can never be shadowed

Target branch

• Base is develop (feature / fix — default)

Type of change

• Bug fix

Test plan

• npx tsc -b — clean

• npx eslint on both changed files — clean (the two pre-existing react-hooks/exhaustive-deps warnings in ACP/index.tsx are untouched)

• npx vitest run src/pages/Agent/ACP — 3 passed: new ACPPanel.test.tsx covers re-adding a disabled custom runner (single PUT, re-enabled) plus both rejection paths (enabled custom / disabled builtin → toast, no request at all)

• Verified the regression test fails on the unpatched guard: with the old code the "re-adds a disabled custom runner" case times out while runnerKeyExists is shown, and passes with the fix

• Full-suite npx vitest run (local Windows): 1187 passed; the remaining failures are pre-existing on the develop base and unrelated to this diff (same set documented in fix(dashboard): localize remaining hardcoded Chinese UI strings (browser AI panel, skill recording, PWA guides) #1799): DOMMatrix is not defined (pdfjs in jsdom) when collecting docxSanitize / skillMarkdown / SkillDrawer, knowledgeBases / publishedExperts mocks on the base missing the newer bridgeAgentHeaders export, and one rc-notification close-timer error attributed to ChannelsPanel.test.tsx

• Added/updated tests

• make all passes locally (the Python harness make all does not apply to a dashboard-only change; the dashboard checks above were run instead)

Checklist

• Updated CHANGELOG.md (if user-facing)
• README / docs updated (if needed)

[fengyue-xve]

The Windows job is red on the unrelated flaky test_cold_target_install_budget_preserves_regular_timeout (float wobble in the 900s budget assertion); fix proposed in #1826.