Specialization Constants and Floating-Point Operations

Creating this issue from a comment from an OpenCL 3.0 issue:

[#294 (comment)](https://github.com/KhronosGroup/OpenCL-Docs/issues/294#issuecomment-648970076)

> I would note the SPIR-V specification also describes floating-point operations for **OpSpecConstantOp** when using the **Kernel** execution model. It is typical to implement compile-time evaluation of floating-point operations using a software library, rather than relying on the floating-point unit of the system running the compiler (or in this case, the specializer). Otherwise, the precise behavior of the compiler (or specializer/translator) will be dependent on the system upon which it runs, with respect to rounding and the handling of NaNs. Determinism of the compiler (or specilizer/translator) is a desirable quality, and as such, implementing **OpSpecConstantOp** for floating-point operations to a reasonable standard would require adding extra dependencies to the _Khronos SPIR-V/LLVM Translator_, which is not a decision that should be made lightly.

Start with the linked comment in OpenCL issue #294 and the SPIR-V specification sections covering OpSpecConstantOp and floating-point operations under the Kernel execution model. Determine whether the OpenCL documentation needs clarification about deterministic evaluation, rounding, and NaN handling; done means the expected specification behavior is resolved and documented.

Written by the indexing model from the issue text.