feat(windows): add hidden OCR context capture from frontmost window for post-processing

[evrenesat]

Summary

This PR adds Windows support for the hidden OCR context template variable (${OCR} / ${ocr}) used in post-processing prompts.(Windows counter-part of #770)

The captured text is read from the current foreground window and injected only into the prompt payload.

Why

This improves prompt quality for post-processing while keeping the app local-first and extensible:

• Better context-aware rewriting/correction prompts
• No cloud dependency added
• Clear platform-specific implementation that is easy to fork/extend

What changed

• Added Windows OCR capture module:
  • src-tauri/src/windows_ocr.rs
• Wired Windows OCR into prompt-building flow:
  • src-tauri/src/actions.rs
• Registered Windows OCR module:
  • src-tauri/src/lib.rs
• Added required Windows API feature flags:
  • src-tauri/Cargo.toml

Behavior

• OCR context is only used if prompt contains ${OCR} or ${ocr}.
• If OCR capture fails, processing continues with empty OCR context.
• On unsupported platforms for this implementation path, OCR context resolves to empty string.

Testing

Tested on Windows:

1. Configure a post-processing prompt containing ${OCR}.
2. Start transcription with post-processing enabled.
3. Verified prompt receives OCR text from frontmost window.
4. Verified fallback to empty OCR context when capture fails.
5. Verified normal transcription flow is unchanged.

Note: I have built and manually tested this on a Windows 11 ARM guest running on MacOS. Even with help of Codex, it took hours to build on arm64. I tested that post-processing works both with and without OCR prompts. In my limited testing, with OCR, model consistently managed to use correct symbols in dictation result (e.g. createUpdaterArtifacts instead of "create-updater-artifacts" or "Create Updater Artifacts" or "create updater artifacts").

Scope / Non-goals

• This PR is Windows-only OCR.
• No local machine build-workaround changes are included in this PR. (hopefully)
• No updater/signing config changes included.

Breaking changes

None.

AI Assistance Disclosure

• AI used: Yes
• Tools used: GPT-5 Codex / ChatGPT
• Extent: Assisted with implementation scaffolding, conflict resolution, and debugging workflow; final code reviewed and validated manually.

Note: I'm planning to try to create another PR for Linux as well, using Tesseract.

[cjpais]

@evrenesat would you mind merging this into the same PR as #770? Would rather have the feature come in all at once when I decide to merge it. Closing this PR for now, I'm trying to slim down the number of active PR's so I can have some sanity and manage maintenance.