feat: scaffold Fish Audio speech provider plugin

- index.ts: plugin entry with definePluginEntry + registerSpeechProvider
- speech-provider.ts: full SpeechProviderPlugin implementation
  - resolveConfig from messages.tts.providers.fish-audio
  - parseDirectiveToken for voice, model, speed, latency, temperature, top_p
  - listVoices merging official + user's own voices
  - synthesize with format-aware output (opus for voice-note, mp3 otherwise)
  - stub Talk Mode (resolveTalkConfig/resolveTalkOverrides)
- tts.ts: raw fishAudioTTS() fetch + listFishAudioVoices()
  - streaming chunked → buffer, error body included in exceptions
  - parallel voice listing with graceful partial failure
- speech-provider.test.ts: voice ID validation tests
- openclaw.plugin.json: speechProviders contract
- package.json: peer dep on openclaw >=2026.3.0

openclaw.plugin.json

@@ -0,0 +1,11 @@
+{
+  "id": "fish-audio",
+  "contracts": {
+    "speechProviders": ["fish-audio"]
+  },
+  "configSchema": {
+    "type": "object",
+    "additionalProperties": false,
+    "properties": {}
+  }
+}