[mlir][dxsa] Document well-defined out-of-range / NaN behavior of float-to-int conversions

[tagolog]

The float/double to integer conversion ops (ftoi, ftou, dtoi, dtou) have a behavior for inputs that fall outside the representable range of the destination type, and for NaN inputs:

• out-of-range inputs are clamped to the destination type's range ([INT32_MIN, INT32_MAX] for the signed *toi ops, [0, UINT32_MAX] for the unsigned *tou ops);
• NaN converts to 0.

Currently none of the four op descriptions in DXSATypeConversionOps.td mention this.
We need to refine proper description for the operations.

Discussion: PR #187 (see @hvdijk's review comment on the dtou description).