fix(speech-to-text): correct output storage path

[IzaakGough]

Summary

Fixes #754

• derive the transcoded upload destination from the source object name instead of the local tmp path
• normalize OUTPUT_STORAGE_PATH before constructing the destination key
• add regression tests covering nested input paths and matching output prefixes

Testing

npm test -- --runInBand util.test.ts

npm run build

Manual verification

• deployed the extension and uploaded myDirectory/test-audio.m4a
• verified outputs were created under OUTPUT_STORAGE_PATH/ without nested source folders or tmp path leakage

[IzaakGough]

Note that failing checks are unrelated to this PR