fix(i18n): keep production sourcemaps usable when the i18n plugin rewrites a chunk

[totally-not-ai]

Summary

The Vaadin i18n build plugin replaced the sourcemap of every chunk with an empty one, so all .map files of an app built with build.sourcemap enabled were useless. The plugin now rewrites chunks at a point where the bundler keeps the original sourcemap, and leaves untouched chunks alone.

Related to #16679

What changed

Behavior change: only apps that build with build.sourcemap enabled are affected, and only for the better — their .map files now contain the real sources instead of being empty. Chunk output itself is unchanged, so no API or runtime behavior changes for anyone else.

• rollup-plugin-vaadin-i18n.js moves the chunk rewriting from generateBundle to renderChunk. The bundler then composes the map returned by the plugin with the map it already has for that chunk. Chunks the plugin does not touch return null and keep their sourcemap as is. generateBundle still collects the translation keys.
• The registerChunk calls are now found with a regular expression instead of an exact string match. It tolerates a renamed i18n binding and rewritten string quotes, so duplicate calls are removed and the chunk name marker is replaced also when the bundler has reformatted the rendered chunk.
• The vite-production test app now uses Hilla translations in two frontend modules (via a small stub of @vaadin/hilla-react-i18n) and builds with sourcemaps on, so the tests actually run through the rewriting path.

Test summary

#	Status	What the test verifies	Why it matters
1	✅	Every .map file reachable from the entry bundle has non-empty sources, non-empty mappings, and the contents of each source	This is the bug: a rewritten chunk got a map with no sources, so the browser could not map the bundle back to the original files
2	✅	Chunks the plugin does not rewrite still have a usable sourcemap	The early return null must not drop maps the bundler produced
3	✅	The chunk that contains translations calls registerChunk with its own served file name, and the marker is gone	A wrong or unreplaced name means the runtime cannot load the translations of that chunk
4	✅	That chunk has exactly one registerChunk call	Duplicate calls from several modules must be removed; a missed match would also leave a stale marker behind
5	❗ gap	Matching when the bundler renames the i18n binding or rewrites the string quotes	The regex exists for this case, but no test forces that output shape; whether it happens depends on the minifier settings of the test app

Tests added on this branch:

• SourceMapsIT.bundleSourceMapsPointToOriginalSources — rows 1, 2
• I18nChunkIT.chunkRegistersItselfUnderItsOwnName — rows 3, 4

Left untested on purpose: BundleAccess is a test helper that only downloads bundle files, so it is covered indirectly by both ITs.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[github-actions]

Test Results

 1 442 files  +2   1 526 suites  +2   1h 31m 36s ⏱️ - 3m 1s
12 062 tests +2  11 994 ✅ +2  68 💤 ±0  0 ❌ ±0 
12 380 runs  +2  12 312 ✅ +2  68 💤 ±0  0 ❌ ±0 

Results for commit 78f3383. ± Comparison against base commit 980ac5b.