feat: add Windows STT support with cross-platform process management and configurable whisper path

[crixue]

feat: add Windows STT support with cross-platform process management and configurable whisper path

It Require merge feat: enhance audio backend detection for Linux/WLS2 and input devi…
first!

Why
The plugin was Linux/macOS-only for STT. Windows users had no supported path
to use speech-to-text despite sox and whisper.cpp both being available on
the platform. The existing code also used Unix-only signals (SIGKILL/SIGINT)
and utilities (pkill) that crash on Windows.

What Changes

• Guard Unix-specific process.kill(pid, "SIGKILL"/"SIGINT") and
  pkill calls behind process.platform !== "win32"; use .kill() on
  Windows where signal names are unsupported
• Add whisperPath plugin option so users can point to the binary directly,
  bypassing PATH resolution — essential on Windows where OpenCode's process
  environment may predate PATH modifications
• Improve ENOENT error message to suggest whisperPath as a fix
• Document full Windows setup in README: winget for sox, Visual Studio Build
  Tools, CPU-only and CUDA whisper.cpp builds, PATH configuration, and smoke
  test
• Generalize README CUDA docs (CUDA_ARCH env var, Blackwell min CUDA 12.8,
  -real suffix for Windows CUDA arch flags)

Capabilities

• STT recording and transcription works on Windows with the same tiny →
  large-v3 model range as Linux/macOS
• NVIDIA GPU acceleration on Windows (CUDA build)
• Users can avoid PATH issues entirely by setting whisperPath in plugin
  options

Impact

• Zero behavioral change on Linux/macOS — all conditional branches are
  no-ops on those platforms
• Windows users get a complete, documented setup flow from winget install to
  working voice commands
• whisperPath option is additive and optional; existing configurations
  continue to resolve via PATH as before
• CUDA arch table fix (Blackwell min 12.8) corrects an error for all platforms