Enable customizable logging backends and logger subclassing

[MarkusPorti]

Implements changes proposed / discussed in #207

Summary

• Introduces LogWriter, an ABC that defines the interface for logging backends.
• Logger now accepts an optional writer: LogWriter constructor argument for injecting a custom backend (e.g. MLflow, a database) without touching the config.
• OnPolicyRunner.__init__ adds Logger as possible type for log_dir parameter. Pass a path string for default behavior, or a pre-built Logger (or subclass) instance for full customization.
• OnPolicyRunner.__init__ gets a new parameter writer. Pass a custom LogWriter to choose where the default logs are stored.
• WandbSummaryWriter and NeptuneSummaryWriter now explicitly inherit from LogWriter, replacing ad-hoc logger_type string guards with isinstance checks throughout Logger.
• LogWriter and Logger are exported from rsl_rl.utils.

Extension patterns

Custom backend (where data goes) — subclass LogWriter:

from rsl_rl.utils import LogWriter, Logger

class MLflowWriter(LogWriter):
    def add_scalar(self, tag, scalar_value, global_step):
        mlflow.log_metric(tag.replace("/", "."), scalar_value, step=global_step)
    def stop(self):
        mlflow.end_run()

runner = OnPolicyRunner(env, cfg, writer=MLflowWriter())

Custom metrics (what gets logged) — subclass Logger:

from rsl_rl.utils import Logger

class MyLogger(Logger):
    def log(self, it, start_it, total_it, ...):
        super().log(it, start_it, total_it, ...)
        self.writer.add_scalar("Custom/my_metric", compute_x(), it)

runner = OnPolicyRunner(env, cfg, log_dir=MyLogger(log_dir="/path", cfg=cfg, ...))

Backward compatibility

All existing usage (cfg["logger"] = "tensorboard"/"wandb"/"neptune", passing log_dir as a string) is unchanged.

[ClemensSchwarke]

Hi @MarkusPorti, thanks a lot for the PR! After reviewing it, I decided to go for a slightly different implementation in #211. I would really appreciate a quick review if you find the time :)

Cheers, Clemens