[feature][refactor] Optimum-Benchmark API

[IlyasMoutawwakil]

In order to support an API that's very simple to use and doesn't require passing by hydra CLI and subprocess.
This refactoring removes most of hydra magic, especially resolvers that made the usage of optimum-benchmark as an API impossible.

Example:

from pprint import pprint
from optimum_benchmark.logging_utils import setup_logging
from optimum_benchmark.experiment import launch, ExperimentConfig
from optimum_benchmark.backends.pytorch.config import PyTorchConfig
from optimum_benchmark.launchers.process.config import ProcessConfig
from optimum_benchmark.benchmarks.inference.config import InferenceConfig

if __name__ == "__main__":
    setup_logging(level="INFO")
    backend_config = PyTorchConfig(model="gpt2", no_weights=True, device="cuda")
    launcher_config = ProcessConfig(device_isolation=True)
    benchmark_config = InferenceConfig(memory=True)
    experiment_config = ExperimentConfig(
        experiment_name="python-api-launch-experiment",
        benchmark=benchmark_config,
        launcher=launcher_config,
        backend=backend_config,
    )
    benchmark_report = launch(experiment_config)
    pprint(benchmark_report)

[2024-02-06 12:58:42,933][backend][WARNING] - Multiple GPUs detected but CUDA_VISIBLE_DEVICES is not 
[...]
[2024-02-06 12:59:14,971][isolation][INFO] -    + Closing device(s) isolation process...
{'decode.latency(s)': 0.7354095666084248,
 'decode.throughput(tokens/s)': 269.2377268263452,
 'forward.latency(s)': 0.0073284958915750915,
 'forward.max_memory_allocated(MB)': 530,
 'forward.max_memory_reserved(MB)': 574,
 'forward.max_memory_used(MB)': 2025,
 'forward.peak_memory(MB)': 2025,
 'forward.throughput(samples/s)': 272.9072963388325,
 'generate.latency(s)': 0.7427380625,
 'generate.max_memory_allocated(MB)': 555,
 'generate.max_memory_reserved(MB)': 618,
 'generate.max_memory_used(MB)': 2069,
 'generate.peak_memory(MB)': 2069,
 'generate.throughput(tokens/s)': 269.2739339718436}

model, device, task and library are deprecated in experiment, in favor of backend.model and backend.device, backend.task and backend.library.
No breaking changes, only more flexibility and fine grained testing 😊.