feat(cdp): honor Browser.grantPermissions / setPermission / resetPermissions

[rohitsux]

Browser.grantPermissions, Browser.setPermission and Browser.resetPermissions were noops, so a CDP automation client (Puppeteer's browserContext.grantPermissions(['geolocation']), Playwright's context.grantPermissions) could not influence what navigator.permissions.query() returns. The query always reported "prompt" regardless. This is a gap for automation flows that need to pre-grant or deny permissions (geolocation, notifications, clipboard, etc.) before a page runs.

Changes

• src/browser/Page.zig: store permission state on the Page, keyed by permission name, with a small setPermission helper. Keys and values are duped into the Page arena, so the state lives for (and resets with) the Page.
• src/cdp/domains/browser.zig: implement the three methods. grantPermissions sets each listed permission to "granted"; setPermission sets a single permission to an explicit "granted"/"denied"/"prompt" state; resetPermissions clears the map.
• src/browser/webapi/Permissions.zig: navigator.permissions.query() now reads the stored state and falls back to "prompt" when unset (previously hardcoded "prompt").

Scope note

Permission state is stored on the active Page, so it resets on navigation. CDP scopes these methods to a browser context (state persists across navigations); persisting across navigations would mean storing on the CDP BrowserContext and re-applying on each new Page, which I left as a follow-up to keep this change small. The origin / browserContextId parameters are accepted and ignored for now.

Verified

• make test passes (796/796, no leaks); new test cdp.browser: grant/set/reset permissions reach navigator.permissions drives all three methods through processMessage and asserts the stored state.
• zig fmt --check ./*.zig ./**/*.zig clean.
• make build succeeds.

[krichprollsch]

Shouldn't we set permissions at the browser's level?
Here, any permission set via CDP will be lost with the page.

Also regarding the origin support, if we don't support it right now, we should at least write down a not_implemented log.
Same about browserContextId.

[rohitsux]

Thanks for the review. Addressed all three points:

• Moved the permission store from the active `Page` to the `Browser` (mirroring how `http_client` holds the user-agent override), so permissions now persist across page navigations instead of being lost with the page. `navigator.permissions.query()` reads the browser-level state via `frame._session.browser`.
• Added `not_implemented` log warnings for `origin` and `browserContextId` on both `grantPermissions` and `setPermission`, since they are accepted but not yet honored.

The test now asserts the state on the browser. `make test` passes (796/796, no leaks), `zig fmt --check` is clean, and `make build` succeeds.

[rohitsux]

Rebased the browser-level change on top of your PermissionState enum refactor (it had gotten clobbered in an earlier merge on my end). The browser store is now StringHashMapUnmanaged(PermissionState), setPermission takes the enum, and Browser.setPermission from CDP parses the setting via std.meta.stringToEnum. Force-pushed. make test is green (796/796, no leaks).