Monomorphize Routine

[DaniPopes]

Currently, calling {Criterion,BenchmarkGroup}::bench_function creates a copy of BenchmarkGroup::run_bench, which includes all the Routine trait methods, for every unique bench closure. This is unnecessary and doesn't make a difference for benchmark performance as the closure in question is the one taking &mut Bencher, which is not the actual routine that is timed.

This PR roughly halves the number of LLVM lines generated by the Rust compiler, tested with recmo/uint's benchmarking suite using cargo llvm-lines -p ruint --bench bench_uint. This suite has around 214 unique calls to criterion.bench_function. This speeds up compilation of a release build of the benchmark from ~15s to ~12s on my machine, likely a lot more in more resource-constrained environments like in CI, or when using less codegen-units.

Before:

  Lines                 Copies               Function name
  -----                 ------               -------------
  911857                26592                (TOTAL)

After:

  Lines                 Copies               Function name
  -----                 ------               -------------
  413095                15305                (TOTAL)