Each fused op runs both transforms in a single loop, eliminating the intermediate buffer. These are NOT `Op` implementors; they are internal helpers used by the fusion picker in `chain::runtime`. # Fused pairs - [`FusedIeee16ByteSplit2`] — `BitReorderIeee16` + `ByteSplit{n=2}` - [`FusedIeee32ByteSplit4`] — `BitReorderIeee32` + `ByteSplit{n=4}` - [`FusedFp8E4M3NibbleSplit`] — `BitReorderFp8E4M3` + `NibbleSplit` - [`FusedFp8E5M2NibbleSplit`] — `BitReorderFp8E5M2` + `NibbleSplit`

For each FP8-E4M3FN byte: reorders bits so the exponent is in the high nibble, then splits into high-nibble (exponent) and low-nibble (sign+mantissa) planes — all in one pass.

For each FP8-E5M2 byte: reorders bits so the exponent is in the high nibble, then splits into high-nibble (exponent) and low-nibble (sign+mantissa) planes — all in one pass.

In `forward`: applies IEEE-754 16-bit float bit reordering in-place, then splits even/odd bytes into two planes — all in a single pass. In `inverse`: merges two planes and reverts the bit reordering.

Applies IEEE-754 32-bit float bit reordering, then splits bytes into 4 planes by round-robin — all in one pass.