Add Benchmarks

[PhilippNaused]

Add a simple benchmark project.
It tests the performance of the tokenizer, text pre-processing and inference.
You can run it by calling:
dotnet run --project Benchmark -c Release

Example output:

| Method         |         Mean |       Error |      StdDev |
|----------------|-------------:|------------:|------------:|
| PreprocessText |     202.9 μs |     1.13 μs |     1.05 μs |
| Phonemize      | 148,545.9 μs | 2,866.41 μs | 2,541.00 μs |
| Tokenize       | 147,700.7 μs | 2,047.23 μs | 1,814.82 μs |

| Method | Model   |       Mean |    Error |   StdDev |
|--------|---------|-----------:|---------:|---------:|
| CPU    | float32 |   462.5 ms |  9.19 ms | 16.33 ms |
| CUDA   | float32 |   134.8 ms |  3.20 ms |  9.43 ms |
| CPU    | float16 |   432.3 ms |  8.63 ms | 15.99 ms |
| CUDA   | float16 |   214.0 ms |  3.36 ms |  3.14 ms |
| CPU    | int8    | 1,887.4 ms | 37.52 ms | 86.21 ms |
| CUDA   | int8    | 2,106.5 ms | 41.13 ms | 57.66 ms |

[PhilippNaused]

@Lyrcaxis Could you review this when you have the time?

[Lyrcaxis]

@PhilippNaused Hey, thanks for the PR! I've reviewed this for some time now -- code looks good.

I've been thinking on whether we want a benchmark project in the repo -- what value it would add, any pitfalls, etc.
Thoughts so far:

Pros: I guess it could get curious people somewhat more involved to making edits and comparing performance (?)
Cons: It might not be very relevant past the GPU vs CPU benchmarks. Not until batching is officially supported, at least.

So I thought I'd give it some time since it's not an engine feature, but sorry, it's indeed been a while 😅 (month+)

Are you looking forward to using the Benchmark project for more stuff/automations/metrics/docs / expand on it?
In that case it can get merged asap. Anyway let me know your thoughts.

[Lyrcaxis]

Cool! Few comments and minor styling tweaks here and there.

Let's talk about the potential of this benchmark project for a bit first before merging though.

[Lyrcaxis]

That would require the benchmarker to tweak a non-highest-level file if they don't use CUDA.
Preferrably this would be an option configurable from the highest-level Program.cs or the CLI args.

[PhilippNaused]

I think the best option would be to split the benchmark into two classes. One for comparing the models on CPU and one for CPU vs GPU.
e.g.,

public class Inference {
...
}

public class InferenceCPUvsCUDA : Inference {
...
}

Then you can select which one you like from the CLI using BenchmarkDotNet built-in benchmark selector.

I'm just not too happy with the name InferenceCPUvsCUDA...

[Lyrcaxis]

I'm not a fan of implicit Program/Main but could make sense here if it's a one-liner.
I think it shouldn't be a one-liner though and could have more options.

[PhilippNaused]

Do you mean you would prefer this?

BenchmarkDotNet.Running.BenchmarkSwitcher.FromAssembly(typeof(Program).Assembly).Run(args);