fix: show days for long ETA durations

[masahiro-kubota]

Description

Fix duration formatting in training logs when elapsed time or ETA exceeds 24 hours. The previous %H:%M:%S formatting drops the day count, so an ETA around 2 days 22 hours can appear as only 22:xx:xx.

Reproduction

From this branch:

python - <<'PY'
import importlib.util
import pathlib
import torch

logger_path = pathlib.Path("rsl_rl/utils/logger.py")
spec = importlib.util.spec_from_file_location("rsl_rl_logger", logger_path)
logger_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(logger_module)
Logger = logger_module.Logger

class DummyWriter:
    def add_scalar(self, *args, **kwargs):
        pass

logger = Logger(
    log_dir=".",
    cfg={"algorithm": {"rnd_cfg": None}, "num_steps_per_env": 24},
    env_cfg={},
    num_envs=10240,
    is_distributed=False,
    gpu_world_size=1,
    gpu_global_rank=0,
    device="cpu",
)
logger.writer = DummyWriter()
logger.logger_type = "tensorboard"
logger.tot_time = 7 * 3600 + 2 * 60 + 10.427
logger.log(
    it=4540,
    start_it=0,
    total_it=50000,
    collect_time=5.287,
    learn_time=0.286,
    loss_dict={"value_function": 13.0612},
    learning_rate=0.0,
    action_std=torch.tensor([0.29]),
    rnd_weight=None,
    print_minimal=True,
)
PY

Test performed

• PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest tests/utils/test_logger.py -q
• Ran the reproduction command above after the fix and confirmed the ETA prints as 2 days, 22:27:13 instead of wrapping to 22:27:13.