stream : add interactive start/stop mode (-i)

[ailakshya]

Closes #969.

What

Adds an -i / --interactive flag to the stream example. Instead of transcribing continuously from launch, stream starts paused; the user presses SPACE to start/stop microphone capture and Q to quit.

Why

The request in #969 — "press space bar to start and stop recording … rather than [it doing] automatic speech recognition." The always-on default is awkward for real dictation/push-to-talk: you transcribe ambient audio and can't cleanly delimit an utterance. It's also related to the duplicate-lines discussion in #1056 / #3866 — several users just want the audio they chose to record.

How

• New -i / --interactive flag — default off, so existing behavior is unchanged.
• Minimal non-blocking keyboard read: termios (raw, no-echo) on POSIX, conio (_kbhit/_getch) on Windows. Guarded by isatty(STDIN_FILENO) → no-ops on pipes/redirects, so scripted use is unaffected.
• Terminal state is saved and restored on exit and via atexit(), so the shell is never left in raw mode (even on Ctrl+C).
• In interactive mode capture starts paused (audio.pause()); SPACE toggles audio.resume()/pause() (with audio.clear() on start to drop buffered silence); while paused the loop sleeps and skips inference. Works in both sliding-window and VAD (--step 0) modes.

Scope

• 1 file, +76 / −1. No new dependencies (SDL already present; termios is POSIX-standard).
• Deliberately terminal-based rather than SDL keyboard: stream opens only an SDL audio device (no video window), so SDL key events aren't delivered — stdin is the right channel for a console app.

Testing

• Compile-tested on Linux (CPU and Vulkan builds).
• Interactive flow verified end-to-end (SPACE → recording, SPACE → stopped, Q → clean exit, status 0), driven through a PTY with injected keystrokes.

Usage

./build/bin/whisper-stream -m models/ggml-base.en.bin -i
# [Interactive mode: press SPACE to start/stop recording, Q to quit]