YeoJohnsonTransformer inverse_transform loses small nonzero values #1100

Describe the bug
YeoJohnsonTransformer.inverse_transform loses nonzero values near zero through cancellation in the inverse formulas. This reproduces on Feature-engine 1.9.4 and main d866312874a66364c6dee3c247e69476b478395c.

To Reproduce

import pandas as pd
from feature_engine.transformation import YeoJohnsonTransformer

transformer = YeoJohnsonTransformer().fit(
    pd.DataFrame({"x": [-3.0, -1.0, 0.0, 1.0, 3.0]})
)
x = pd.DataFrame({"x": [-1e-20, 0.0, 1e-20]})
transformed = transformer.transform(x)
print(transformed["x"].tolist())
print(transformer.inverse_transform(transformed)["x"].tolist())

The output is [-1e-20, 0.0, 1e-20] followed by [0.0, 0.0, 0.0]. The fitted lambda is approximately 1; no fitted attribute needs to be overridden. Normal-sized values such as [-1.0, 0.0, 1.0] round-trip correctly.

Expected behavior
The inverse should recover the nonzero values with relative accuracy, as the forward transformation does. The 1 + x and subtraction of 1 in the inverse formulas round away these values.

Desktop
Windows, Python 3.12.14, NumPy 2.2.6, pandas 2.3.3, SciPy 1.16.3.

Additional context
I am working on a small fix using log1p and expm1 for the mathematically equivalent inverse formulas, with near-zero round-trip regressions and normal-sized controls. This retains the inverse behavior introduced in #679/#696 and the index correction in #794. The open narwhals migration #965 retains the current cancellation-prone formulas, so this is a separate numerical correction.