Code review fixes for float math flash savings

Follow-up to the single-precision math change that saves ~17 KB of
flash on STM32F722 targets:

- maths.c: Replace powf(x, 2.0f) with explicit multiplies in
  gaussian(). Clearer intent and avoids any dependency on libm
  reducing integer-exponent powf to a multiply.

- vtx_smartaudio.c: Add roundf() before uint16_t cast in saDbiToMw().
  powf(10.0f, 1.0f) may return 9.999...f on some libm implementations,
  which would truncate to 9 mW instead of the correct 10 mW. roundf()
  closes the hazard at the cost of one FPU instruction.

Author: sensei-hacker

src/main/common/maths.c

@@ -526,7 +526,7 @@ bool sensorCalibrationSolveForScale(sensorCalibrationState_t * state, float resu
 }
 
 float gaussian(const float x, const float mu, const float sigma) {
-    return expf(-powf(x - mu, 2.0f) / (2.0f * powf(sigma, 2.0f)));
+    return expf(-((x - mu) * (x - mu)) / (2.0f * sigma * sigma));
 }
 
 float bellCurve(const float x, const float curveWidth)

src/main/io/vtx_smartaudio.c

@@ -212,7 +212,7 @@ int saDacToPowerIndex(int dac)
 
 int saDbiToMw(uint16_t dbi) {
 
-    uint16_t mw = (uint16_t)powf(10.0f, (float)dbi / 10.0f);
+    uint16_t mw = (uint16_t)roundf(powf(10.0f, (float)dbi / 10.0f));
 
     if (dbi > 14) {
         // For powers greater than 25mW round up to a multiple of 50 to match expectations