perf: reduce perceived recording start delay by ~50%

[JereSalo]

Summary

Reduces the perceived delay between pressing the hotkey and seeing the recording overlay by ~50%, from ~850ms to ~400-450ms. Two independent optimizations:

1. Show overlay in parallel with engine startup (1c2ca54)

Previously, the overlay expand animation only started after the audio engine was fully initialized, because it was triggered by isRunning becoming true. This adds showOverlayEarly() which fires the overlay immediately on hotkey press, so the ~300ms expand animation runs concurrently with the ~330ms engine startup instead of sequentially after it. The overlay doesn't need the engine to be ready — it only needs the audioLevelPublisher, which simply emits no values until the engine starts capturing. Applied to dictation, command, and rewrite mode callbacks.

2. Keep audio engine alive between recordings (dd2c73d)

Previously, stop() destroyed the AVAudioEngine instance and the next start() had to recreate it from scratch, including the expensive inputNode (~166ms) and outputNode (~36ms) instantiation where CoreAudio creates the aggregate device. Now stop() just calls engine.stop() and keeps the instance in memory. On the next start(), the nodes are already instantiated so configureSession() is near-instant, and only prepare() + start() need to run. Device changes between recordings are handled normally because in sync mode the engine follows the system default device, and prepare() re-negotiates the audio format with whatever device is currently active.

Measured results

Engine startup time (hotkey callback to isRunning = true):

Scenario	Time
Original v1.5.9	~360ms
First recording after launch	~130-190ms
Subsequent recordings (engine warm)	85-140ms

Total perceived delay (Option key press to overlay fully visible):

Scenario	Time
Original v1.5.9	~850ms
With both optimizations	~400-450ms

[JereSalo]

Hi! I made this PR just because I upgraded to 1.5.9 and though it's fast once you finish speaking at the beginning I found it kinda sluggish, taking almost a whole second from when I pressed option and I saw the animation to start talking, so I made a few changes regarding this, the most important one is initiating the UI and the engine concurrently so that we don't need to wait for the engine to load till we load the UI. This works fine because the UI initialization is not very fast.
I'll be using this version.

[altic-dev]

@JereSalo thanks! please do use it for a couple days and lmk how it goes. The issue is it might work for you and i but will cause issues for others with different devices etc. so I made that. delay to be safe. But down to check this out once you think it's stable :)