fix(foundry): keep local ASR warm and stream transcripts

[H-Chris233]

Summary

• expose the Foundry keep-loaded preference and interpret the existing 86400 value as no automatic unload across local ASR runtimes
• cache successful Foundry execution-provider registration once per process so model reloads do not repeat catalog initialization
• collect Foundry streaming transcription responses in order instead of using the one-shot path that reproduced U+FFFD corruption
• update the shared contracts, Windows UI, neutral local-ASR copy, and focused regression coverage

Related to #1031.

Validation

• npm.cmd test (83 frontend tests and production build)
• cargo test -p openless-core --locked (794 passed, 1 ignored, plus integration suites)
• cargo test --lib for the Windows Tauri crate (487 passed, 1 ignored)
• focused streaming collector tests (2 passed)
• focused forever-residency test (1 passed)
• targeted rustfmt --check
• git diff --check

Remaining acceptance

• The five-minute real-app Foundry residency/catalog-log check was not run against this build.
• The ten-run real-app U+FFFD check was not run against this build.
• This does not claim to fix the separate long-utterance missing-words report; [windows][foundry-local] 本地 Whisper 空闲卸载后冷启动很慢；长句截断且末尾变成 �；希望长驻模式与本地润色模型 #1031 should remain open pending the reporter's original WAV.