Optimize log message code generation

[sethrj]

While testing the contents of g4vg.dylib with nm to make sure no Celeritas physics was included, I found that template instantiations for the logger message on char const[N] were being saved as actual function calls. This is pretty silly because those functions unroll effectively to put_char(ptr, N), so there's not going to be any code size gain by not inlining those.

The second optimization is that I've marked as "unlikely" that the log level is high enough to print, moving the printing code to a "cold" section. This improves instruction cache locality in the typical case where a logger message is ignored.

I've also replaced std::string with std::string_view and made better use of move semantics to further reduce the amount of code needed to create and destroy a logging message, especially when it's not printed.

---

For reference, given the code:

void write_message(int i, char const* s)
{
    CELER_LOG(debug) << "Error: expected " << i << " to be " << s
                     << "(but double is " << 1.234 << ')';
}

With the new code, there are 15 instructions between "get the world logger" and the function exit. With the old code, there's ~120 (but the "not logging" case skips some of those). The old code also has several calls to delete due to using string rather than string_view.

[pcanal]

LGTM