Fixes for double precision erfc.

[hvdijk]

Overview

Fixes for double precision erfc.

Reason for change

Describe how the current behaviour of the project is causing problems for you
or is otherwise unsatisfactory for your use case.

Description of change

• polynomial5 did not provide as many coefficients as needed for sufficiently accurate results around x=7. Use sollya to recompute.
• s6, s7, s8 multiplied by x rather than by xAbs, leading to incorrect results for x<0.
• s6, s7, s8 lost precision in the multiplication by x. Use multiply_exact_unsafe instead, and then transform exp(mul_hi + mul_lo) into exp(mul_hi) * exp(mul_lo).
• The transformation of erfc(-x) = 2 - erfc(x) checked the sign bit of the intermediate result, rather than the sign bit of x, again leading to incorrect results for x<0.
• The cut-off point for returning 0 was set as x > 27.0, but this is not right, the cut-off point should be slightly higher.

Anything else we should know?

These errors were caught by a new test in OpenCL CTS.

Checklist

• Read and follow the project Code of Conduct.
• Make sure the project builds successfully with your changes.
• Run relevant testing locally to avoid regressions.
• Run clang-format-17 on all modified code.