fix bug where multiple utility code dependencies could be ignored due to loop variable reassignment (only last one wins)

use tuple instead of list for f-string joining because it is faster to allocate - speedup can be up to 20%

we actually know that FormattedValueNode always returns a Unicode string and no other object type => set node type accordingly

make sure f-string formatting always returns a Unicode string (even when not joining the final result in trivial cases)

fix repeated expression evaluation in f-string implementation for !s/r/a conversions by switching from macro to inline function

The cPython profiler would get confused when the wrapper and
underlying cdef function shared a name.