Slightly optimize CPU util tracking by defer float conversion

Previously _onSample() computed and stored a volatile float _sma_avg
on every sample (64 times/minute on core 0).  The getter simply returned
that pre-computed value.

This commit inverts the responsibility:

- _sma_sum is now volatile uint16_t — the raw integer sum is the
  cross-core shared state, written atomically by core 0 (single S16I
  instruction on Xtensa LX7) and read by core 1.

- The float conversion (uint16_t → float multiply) is moved into
  getCore0Util(), which is called rarely (once/minute for history,
  on-demand for web requests).  The multiply now happens on the core
  that actually needs the result.

- _sma_avg is removed entirely, saving 4 bytes and one volatile float
  store per sample from the hot path.

- The SMA update is rewritten as a single combined expression:
    const uint16_t last = _sma_buf[_sma_idx];
    _sma_buf[_sma_idx] = s8;
    _sma_sum = (_sma_sum - last) + s8;
  This produces a single write to _sma_sum instead of two (decrement
  then increment), which is cleaner when _sma_sum is volatile.

The cross-core contract is unchanged: core 0 writes _sma_sum once per
sample; core 1 reads it in getCore0Util().  A 16-bit aligned store on
Xtensa LX7 is a single instruction, so no spinlock is needed.

arch/esp32/CPUUsageTracker.h

@@ -11,7 +11,9 @@ class CPUUsageTracker {
 public:
   void begin();
 
-  float getCore0Util() const { return _sma_avg; }
+  float getCore0Util() const {
+    return (float)_sma_sum * (1.0f / (SMA_WINDOW * 255.0f));
+  }
 
 private:
   static constexpr uint8_t SMA_WINDOW = 64;  // power of 2 — enables & mask
@@ -27,10 +29,9 @@ private:
   uint32_t _last_busy = 0;
 
   uint8_t  _sma_buf[SMA_WINDOW] = {};
-  uint16_t _sma_sum = 0;
   uint8_t  _sma_idx = 0;
 
-  volatile float _sma_avg = 0.0f;
+  volatile uint16_t _sma_sum = 0;
 
   esp_timer_handle_t _timer = nullptr;