chore(web): sync speech recognition language options with SRI 0.2.1

[cswaney]

Summary

The service launcher offered only 8 languages and always sent one on every request. SRI 0.2.1's `TranscriptionRequest` accepts ~90 languages plus per-file auto-detect, so 82 valid values were unreachable and users could never opt into auto-detection.

• Widen the list to mirror the batch transcribe launcher's 12-language curated set: English, Spanish, French, German, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, Hindi, Russian.
• Add "Auto-detect" as the default. `requests.js` now omits the `language` field when Auto-detect is chosen — same mechanism the batch task uses.
• Portal the language dropdown with `anchor="bottom start"` so it escapes the sidebar's overflow context; previously the open menu was clipped by the parameter container's height.

Test plan

• `npm run lint`
• `npm test` (556 pass)
• Manual: launch a speech-recognition service, transcribe an audio file with a specific language and with Auto-detect. Confirm the dropdown is fully visible when open.

Notes

• Sub-task of Update container versions #451.
• The `response_format` toggle (text vs json+segments) is a separate feature request — web: expose response_format toggle for speech-recognition service (text vs json+segments) #521.

Closes #514

[claude]

Review

Nice, focused fix — the auto-detect omission logic and the anchor="bottom start" portal fix both follow patterns already established elsewhere in the codebase (e.g. ModelSelect.jsx, RevisionSelect.jsx), so the dropdown-clipping fix looks solid.

A few things worth a look before merge:

Regression: three previously-supported languages are dropped
The old list included Norwegian, Swedish, and Dutch; the new curated 12-language set (mirroring the batch transcribe launcher) drops them in favor of Italian/Portuguese/Japanese/Korean/Arabic/Hindi/Russian. Since SRI 0.2.1 apparently supports ~90 languages, was narrowing to exactly the batch launcher's set (rather than a larger curated list, or all 90) a deliberate product call? Users who relied on Norwegian/Swedish/Dutch in this launcher lose access to them with no obvious workaround other than picking Auto-detect. Worth confirming this is intentional and not just an incidental side effect of "mirroring" the batch task's list.

Wire-format mismatch vs. the launcher it claims to mirror
The batch transcribe launcher (NewJobModal.jsx) sends ISO 639-1 codes ("en", "es", "pt", …), but this PR's requests.js sends lowercased full names ("english", "spanish", "portuguese", …) via params.language.name.toLowerCase(). That may well be correct — the two code paths likely hit different backends/libraries with different expected formats — but the PR description's "mirror the batch launcher's language set" framing could give a future reader the impression the wire format matches too. Since SRI's source isn't vendored in this repo, this can't be verified statically; the manual test-plan item (checked as [ ]) is the right way to close the loop here, so make sure that actually gets run — including a couple of the newly-added languages (e.g. Portuguese, Hindi), not just English — before merging.

No test coverage for the new/changed behavior

• requests.js's new conditional (if (params.language?.name && ... !== "Auto-detect")) is exactly the kind of small, pure logic that's cheap to unit test and easy to accidentally break later (e.g. reintroducing an unconditional body.language = ...). A test asserting the body omits language for Auto-detect and includes the lowercased name otherwise would be valuable.
• No test exercises the new default (AUTO_DETECT) or the expanded language list in SpeechRecognitionParametersForm.

Minor / pre-existing, not introduced by this PR
LanguageSelect.propTypes still declares language: PropTypes.string, but language is actually an {id, name} object (was already wrong before this change). Not blocking, but could be fixed while this file is already being touched.

Nothing here blocks the core fix — the auto-detect bug and dropdown-clipping bug are real and well addressed. The main ask is confirming the language-list scope-down is intentional and actually running the manual test before merge.