fix: correct subnormal float/double decode exponent

[rophy]

Summary

IEEE 754 subnormal (denormalized) values are decoded with the wrong exponent,
producing approximately half the correct value.

For subnormals (exponent field == 0), the effective exponent should be 1 - bias,
not 0 - bias. The current code falls through without adjusting the exponent,
resulting in values shifted by a factor of 2.

Example

BINARY_FLOAT 1.401298e-45 (smallest positive subnormal) decodes as 7.006492e-46.

Fix

Set exponent = 1 before subtracting the bias in the subnormal path, for both
positive and negative cases of decodeFloat() and decodeDouble().

Test

Standalone test to reproduce and verify:

// Build: g++ -std=c++17 -o test_subnormal test_subnormal.cpp

#include <cmath>
#include <cstdint>
#include <cstdio>
#include <limits>

static double decodeFloat(const uint8_t* data) {
    // ... copy from Builder.cpp ...
}

int main() {
    // Smallest positive subnormal float
    // IEEE 754: sign=0, exp=0x00, sig=0x000001
    // Oracle:   sign flipped -> {0x80, 0x00, 0x00, 0x01}
    const uint8_t data[] = {0x80, 0x00, 0x00, 0x01};
    double result = decodeFloat(data);
    printf("got: %.6e, expected: %.6e\n", result, 1.401298e-45);
    // Without fix: 7.006492e-46 (wrong)
    // With fix:    1.401298e-45 (correct)
}