Reduce string allocations for literals [skip CI]

[bbannier]

With this PR the benchmark from #1589 shows identical performance for both cases for me. Overall runtime seems to improve as well.

Before:

$ hyperfine --export-markdown results.md -w3 'cat test-data.txt | ./build/bin/spicy-driver t.hlto' 'cat test-data.txt | ./build/bin/spicy-driver t-compiled-with-realpath.hlto'
Benchmark 1: cat test-data.txt | ./build/bin/spicy-driver t.hlto
  Time (mean ± σ):      1.713 s ±  0.018 s    [User: 1.699 s, System: 0.031 s]
  Range (min … max):    1.695 s …  1.738 s    10 runs
 
Benchmark 2: cat test-data.txt | ./build/bin/spicy-driver t-compiled-with-realpath.hlto
  Time (mean ± σ):      2.166 s ±  0.014 s    [User: 2.141 s, System: 0.041 s]
  Range (min … max):    2.147 s …  2.191 s    10 runs
 
Summary
  cat test-data.txt | ./build/bin/spicy-driver t.hlto ran
    1.26 ± 0.02 times faster than cat test-data.txt | ./build/bin/spicy-driver t-compiled-with-realpath.hlto

After:

$ hyperfine --export-markdown results.md -w3 'cat test-data.txt | ./build/bin/spicy-driver t.hlto' 'cat test-data.txt | ./build/bin/spicy-driver t-compiled-with-realpath.hlto'
Benchmark 1: cat test-data.txt | ./build/bin/spicy-driver t.hlto
  Time (mean ± σ):      1.218 s ±  0.008 s    [User: 1.204 s, System: 0.029 s]
  Range (min … max):    1.209 s …  1.232 s    10 runs
 
Benchmark 2: cat test-data.txt | ./build/bin/spicy-driver t-compiled-with-realpath.hlto
  Time (mean ± σ):      1.219 s ±  0.005 s    [User: 1.202 s, System: 0.031 s]
  Range (min … max):    1.215 s …  1.231 s    10 runs
 
Summary
  cat test-data.txt | ./build/bin/spicy-driver t.hlto ran
    1.00 ± 0.01 times faster than cat test-data.txt | ./build/bin/spicy-driver t-compiled-with-realpath.hlto

Closes #1589.
Closes #1591.

[bbannier]

The remaining work here is dealing with the fallout from type mismatches from const char* vs std::string. This is likely exactly why we introduced the explicit typing. I picked const char* instead of std::string_view since const char* can implicitly convert to std::string, but maybe introducing a dedicated type would give us more control.

[bbannier]

I slightly reworked this since while we can emit C++ literals for generated (effectively static) strings user-specified literals should still be emitted as mutable strings (e.g., so users can call mutating functions on them). The way I now detect the difference is to check whether a literal has a location.

[bbannier]

I added a couple of other perf cleanups to this PR, please let me know if I should pull them into their own PR.

[rsmmr]

Main comment is that "magic" part: not a fan. Rest looks good.

[rsmmr]

#include <spicy/rt/sink.h>

[bbannier]

Even though we still do not do this consistently, I would really like to include this with #include "..." to make clear that this is the header related to this .cc file. With that we e.g., get it to sort before other headers which helps make sure that the header is self-contained.

[rsmmr]

I think we had this discussion a few times already. :-) I won't fight it, I just don't like the inconsistency of doing it randomly for a few implementation files, but not most others.