[v3] libstdc++/33487

[Benjamin Kosnik]

This fixes up some issues when calling parallel functions.

One issue was with includes: some <algorithm> functions used bits of
<numerics> without an include. Fixed with the include.

As part of this, __gnu_sequential injections were moved to tags.h, and
all of std::__norm in parallel mode was injected into __gnu_sequential
when in parallel mode. This cleans up things nicely.

The major change here is to make sure that there is a parallel mode
function that exactly matches the iso signature. Before, some of the
parallel functions had a compatible signature with an extra argument
(with a default). This was problematic due to the tags being declared
in __gnu_parallel.

tested x86/linux
tested x86/linux parallel

-benjamin