Fix underflow rounding for formats without subnormals

[awf]

round_float and round_ndarray can return values below E8M0's smallest representable value, which subsequently encode as NaN. Clamp finite underflows (including zero) to the smallest magnitude when the target has no zero, preserving sign for signed formats. This applies independently of rounding mode and saturation. This follows CUDA behaviour:
https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/group__CUDA__MATH__FP8__MISC.html

For custom formats that have zero but no subnormals, round across the actual gap between zero and the smallest positive value, using the requested rounding mode. Also fix scalar encoding of exact zero in these formats: select its code directly instead of deriving an exponent from zero.

Add regression coverage for all rounding modes, both saturation settings, stochastic endpoints, positive and negative zero, signed custom formats, tiny inputs, mixed/empty arrays, and the strict Array API backend. Check that rounded values belong to the decoded value set and survive encoding/decoding. Document the lower-clamp policy in both rounding APIs.

Fixes #62.