perf(voice): reduce transmit latency

- FFmpeg: frame_duration 10ms, compression_level 0, nobuffer/low_delay flags, minimal probe
- Frontend: buffer 4096→1024 samples (~170ms→~42ms), faster base64 with chunking
- Remove processor.connect(destination) to avoid audio feedback loop

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

services/discord-gateway/src/modules/voice-recording/transmitter.ts

@@ -50,8 +50,16 @@ export class VoiceTransmitter {
       "-ar", "48000",          // Output sample rate: 48kHz
       "-ac", "2",              // Output channels: stereo
       "-application", "lowdelay", // Low delay mode for real-time
-      "-frame_duration", "20", // 20ms frames
+      "-frame_duration", "10", // 10ms frames (smaller = less latency)
       "-packet_loss", "0",     // No packet loss expected
+      "-compression_level", "0", // Fastest compression (lowest CPU latency)
+      "-cutoff", "20000",      // Audio cutoff frequency
+      "-vsync", "0",           // No video sync
+      "-fflags", "nobuffer",   // No buffering
+      "-flags", "low_delay",   // Low delay processing
+      "-strict", "experimental",
+      "-analyzeduration", "0", // No analysis duration
+      "-probesize", "32",      // Minimal probe size
       "pipe:1",                // Write to stdout
     ], {
       stdio: ["pipe", "pipe", "pipe"],