feat: shared transcription cache to avoid duplicate Whisper API calls

Both services (nanoclaw + nanoclaw-codex) now share cache/transcriptions/
directory. When one service transcribes audio, the other reads from cache
instead of making a second Whisper API call. Uses .pending file to
coordinate concurrent transcription attempts.

src/config.ts

@@ -31,6 +31,8 @@ export const GROUPS_DIR = path.resolve(
 export const DATA_DIR = path.resolve(
   process.env.NANOCLAW_DATA_DIR || path.join(PROJECT_ROOT, 'data'),
 );
+// Shared cache directory (same across both services for dedup)
+export const CACHE_DIR = path.join(PROJECT_ROOT, 'cache');
 
 export const AGENT_TIMEOUT = parseInt(
   process.env.AGENT_TIMEOUT || '1800000',