[carrerasdarren-cell]

Fixes #1077.

optlog2(FLT_MAX) returns 128, which does not fit the signed 8-bit exponent stored by the exponential filter. At low mantissa precision this made optexp2 construct an invalid scale and triggered undefined behavior during float-to-int conversion.

Clamp the stored exponent to 127, scale that boundary without relying on denormal handling, and constrain the mantissa to the largest range that still decodes to a finite float. The existing asymmetric -2^23 mantissa remains available at ordinary exponents.

The regression covers positive and negative FLT_MAX in shared-vector mode, checks the exact encoded words, and verifies finite decoded values.

Tests:

• make -j8 config=sanitize test
• make -j8 config=debug test
• make -j8 config=release test
• make -j8 config=coverage test
• make -j8 config=coverage-scalar test