Support for soft fp16 to fp64 IEEE conversions
This patch adds initial support for lowering conversions from half to
double with single rounding.
No backends have been updated to use this builtin yet.
I'm particularly interested whether the exhaustive testcases make sense for non-identical bittpatterns for NaN values
clang-tidy: warning: invalid case style for function '__extendhfdf2' [readability-identifier-naming]
not useful
clang-tidy: warning: invalid case style for parameter 'a' [readability-identifier-naming]
not useful