Add Linux support: PulseAudio system audio, WebRTC AEC, Vulkan whisper.cpp

[Mathjk]

Summary

Adds Linux support to Wisp, following the existing platform-crate architecture
(wisp-loopback on Windows, wisp-screencapture on macOS):

• crates/wisp-pulseaudio (new): captures the default PulseAudio sink's
  .monitor source on a dedicated capture thread (PulseAudio objects are
  thread-affine), feeding the shared 16 kHz mono f32 frame channel. Works on
  PulseAudio proper and on PipeWire via pipewire-pulse. Includes a readiness
  timeout, a 250 ms heartbeat so next_frame never blocks indefinitely on
  silent output, panic-safe teardown (catch_unwind closes the channel),
  DONT_MOVE so PA module-rescue cannot rebind the stream, and unit tests for
  the pure helpers (frame decoding, sink-monitor naming, wait logic).
• App wiring: open_system_capture() returns the monitor source on Linux;
  wisp-aec (WebRTC AEC3) is enabled for Linux the same way as macOS, with the
  existing passthrough fallback.
• whisper.cpp: the whisper-vulkan cargo feature now also applies to
  Linux, building the vendored ggml Vulkan backend and linking the system
  Vulkan loader. family_runnable / accelerator reporting treat Vulkan as the
  Linux GPU path, matching Windows. Core ML encoder availability is now
  macOS-only (it could never work on Linux builds).
• Packaging: tauri.linux.conf.json bundles the staged sherpa/onnxruntime
  shared libraries plus libvulkan.so.1 under linux-runtime/, declares the
  real runtime Depends: for the deb (webkit2gtk, gtk, pulse, alsa,
  libvulkan1, ...), and bundles deb + appimage targets. build.rs stages the
  runtime libs and links with --disable-new-dtags so the binary gets
  DT_RPATH=$ORIGIN:$ORIGIN/../lib/Wisp (RUNPATH would not propagate to
  transitive deps of the bundled libs).
• CI/release: a linux-app job on ubuntu-22.04 compiles the app in
  every PR (default and whisper-vulkan builds), and the release matrix gains
  an Ubuntu 22.04 leg producing .deb + .AppImage. Two non-obvious toolchain
  details are handled: jammy's archive has no glslc (pulled from LunarG's
  official repo) and its meson 0.61 is older than the vendored
  webrtc-audio-processing meson_version >= 0.63 requirement (installed via
  pip instead).

Verification (Ubuntu 24.04, GNOME/X11, PipeWire via pipewire-pulse)

• cargo check, clippy, fmt clean for the workspace and the new crate;
  8/8 new unit tests pass; 51/51 wisp-models tests pass with
  --all-features.
• Live capture test: silent desktop delivers heartbeat frames (~250 ms
  cadence) without blocking; a 440 Hz tone played through the default sink is
  captured via the .monitor source (measured frames match wall time at
  16 kHz); stop + restart capture works without hangs.
• End-to-end: built .deb + .AppImage, launched the app, downloaded a
  model, recorded a meeting with synthesized speech playing through the
  speakers -- the system-audio monitor path transcribed it in real time and
  attributed it to "them".
• Built artifacts were inspected: correct Depends:, bundled
  sherpa/onnxruntime/vulkan libs, working $ORIGIN-relative RPATH. The
  release build targets Ubuntu 22.04 to keep the glibc floor low; the full
  pipeline (including the vendored ggml Vulkan build and the WebRTC AEC meson
  build) was verified inside an ubuntu:22.04 container.

Notes

• Default-sink changes mid-recording are resolved once at capture start, same
  as the WASAPI loopback behavior.
• The binary also retains ggml's CPU fallback when no Vulkan device is
  present.