Improve documentation for correctly rounded divide and square root

[bashbaug]

We should improve our documentation for the "correctly rounded divide and square root" built option - -cl-fp32-correctly-rounded-divide-sqrt.

Spec reference: https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_API.html#math-intrinsics-options

Some things we should do specifically:

1. Confirm: This build option only applies to single precision divides and square roots, and not to half precision (fp16) divides and square roots?
   • Note that double precision (fp64) divides and square roots are correctly rounded by default.
2. Check: What is the expected behavior if the -cl-fp32-correctly-rounded-divide-sqrt build option is passed to clBuildProgram and the device does not support correctly rounded divides and square roots (CL_DEVICE_SINGLE_FP_CONFIG does not include CL_FP_CORRECTLY_ROUNDED_DIVIDE_SQRT)?
   • Should this generate CL_INVALID_BUILD_OPTIONS?
   • What if some devices support correctly rounded divides and square roots, but some devices do not?
3. Update the math function accuracy tables in the OpenCL C spec and the OpenCL SPIR-V environment spec to clearly document when divides and square roots must be correctly rounded.
   • Low priority: Do we need to define the behavior when both the -cl-fp32-correctly-rounded-divide-sqrt and -cl-unsafe-math-optimizations build options are present?

See also #1499.