Because the carac dsl is implemented in Scala, it is possible to dynamically
compose programs based on information only known at runtime, but most of the
time this level of dynamicity is not necessary because the program is known at
compile-time and only facts need to be loaded at runtime.

We can take advantage of this by leveraging the existing quote backend to
generate programs at compile-time using the standard Scala macro mechanism,
the result is faster than both the interpreter and lambda backend on at least
one simple benchmark:

    BenchMacro.simple_interpreter  thrpt   10  28,511 ± 1,442  ops/s
    BenchMacro.simple_lambda       thrpt   10  27,863 ± 0,397  ops/s
    BenchMacro.simple_macro        thrpt   10  31,917 ± 0,334  ops/s

It'd be interesting to try to extend this system to generate code that can
still be re-optimized at runtime with the JIT.

We still need a small workaround (the macro call needs an explicit `this.`
prefix), but we no longer need to use a different package.

Before:
    Benchmark                          Mode  Cnt  Score   Error  Units
    BenchMacro.ackermann_opt_macro    thrpt   10  1,812 ± 0,019  ops/s
    BenchMacro.ackermann_worst_macro  thrpt   10  0,097 ± 0,002  ops/s
After:
    BenchMacro.ackermann_opt_macro    thrpt   10  1,727 ± 0,042  ops/s
    BenchMacro.ackermann_worst_macro  thrpt   10  1,294 ± 0,036  ops/s

Compute the sort order at runtime and directly use it without an extra
compilation step: instead of re-ordering the code we generate, we re-order the
data the code we generate operates on.

Benchmarks on my unplugged laptop:

Before:
    Benchmark                          Mode   Cnt  Score   Error  Units
    BenchMacro.ackermann_opt_macro     thrpt   10  1,026 ± 0,036  ops/s
    BenchMacro.ackermann_worst_macro   thrpt   10  0,773 ± 0,022  ops/s

After:
    BenchMacro.ackermann_opt_macro     thrpt   10  2,386 ± 0,128  ops/s
    BenchMacro.ackermann_worst_macro   thrpt   10  2,548 ± 0,144  ops/s

This is comparable to results on the lambda backend:
    BenchMacro.ackermann_opt_lambda    thrpt   10  2,556 ± 0,075  ops/s
    BenchMacro.ackermann_worst_lambda  thrpt   10  2,636 ± 0,093  ops/s