fix(audio): fix DMC loop byte skip, add DC blocker, lazy cpal stream

Three audio bugs fixed:

1. DMC loop mode skipped the last byte of each sample iteration.
   provide_dmc_dma_byte() was immediately setting dmc_dma_request on
   loop restart while the sample buffer was still full, causing the
   while-loop in clock_cpu_cycles to service a second DMA immediately
   and overwrite the valid buffer. Per NES hardware spec, the reader
   only fills an empty buffer — the request is now left to clock_dmc
   when the output unit actually empties the buffer into the shift
   register. Fixes intermittent clicking/crackling in games that use
   looped DMC samples (BGM, SFX).

2. Missing DC blocker (high-pass filter) in AudioMixer. The NES APU
   has a capacitor-coupled output stage that blocks DC bias. Without
   it, abrupt channel state changes (length counter expiry, sweep
   mute, triangle period < 2) produce DC steps that manifest as
   audible clicks. Added a one-pole IIR high-pass filter at ~5 Hz
   applied after the existing low-pass filter.

3. cpal stream was opened at application startup with
   BufferSize::Fixed(256), forcing PipeWire/PulseAudio to run the
   entire audio graph at a 5.3 ms quantum. This disrupted other audio
   applications (browsers, media players) even when no ROM was loaded.
   Fixed by: (a) creating the stream lazily on the first push_samples
   call so no device is touched until a ROM is running, and (b)
   switching to BufferSize::Default so the audio server chooses the
   quantum instead of the emulator imposing one. Ring buffer capacity
   increased from 1536 to 4096 samples to absorb larger server quanta.

src/native_core/apu/api.rs

@@ -200,9 +200,15 @@ impl Apu {
         }
         if self.dmc_bytes_remaining == 0 {
             if (self.io[0x10] & 0x40) != 0 {
+                // Loop mode: reset address and byte counter.
+                // Do NOT request another DMA here — the sample buffer is full
+                // right now. clock_dmc will request the next fetch when the
+                // output unit empties the buffer into the shift register, which
+                // is the correct NES hardware behaviour (reader only fills an
+                // empty buffer).  Requesting early would overwrite the valid
+                // buffer and skip the last byte of each loop iteration.
                 self.dmc_bytes_remaining = self.dmc_sample_length_bytes();
                 self.dmc_current_addr = self.dmc_sample_start_addr();
-                self.dmc_dma_request = true;
             } else if self.dmc_irq_enabled {
                 self.dmc_irq_pending = true;
             }