fix(tflm): preflight unsupported int8 Softmax scales

[apage224]

Problem

Some int8 TFLite models reach TFLM PreprocessSoftmaxScaling() with
beta * input_scale * 2^26 <= 1. The shared helper calls
QuantizeMultiplierGreaterThanOne(), which aborts on target during
AllocateTensors() and appears as a HardFault/RTT timeout.

Reproduced with local neuralspot-edge/kws_int8.tflite: final Softmax input
scale 4.305568790385905e-09, multiplier 0.2889418303966522. Both
stock-reference TFLM and Helia-RT/Helia fail because they use the same helper.

Expected

Host-side model preflight detects this unsupported int8 Softmax quantization
range and returns a clear model-compatibility error before firmware generation
or flash.

Notes

The MLPerf-Tiny KWS reference model uses Softmax input scale
0.14469251036643982 and profiles successfully on Apollo510.