Improve loading performance in _absolute_to_relative_idx when remapping indices

[hello3x3]

Type / Scope

• Type: Performance
• Scope: datasets / DatasetReader

Summary / Motivation

This PR improves the initialization performance of DatasetReader when episodes is specified and index remapping is required. Materializing the entire column as Python / tensor objects becomes a significant bottleneck for large datasets. Use .to_numpy() is much faster.

Related issues

• Related: Add absolute_to_reative_idx for remapping indices to fix delta_timestamp bug #2490

What changed

Replaced formatted index column loading with direct Arrow column access in DatasetReader
Optimized remapping dictionary construction for _absolute_to_relative_idx
Preserved existing remapping semantics and public API behavior
No breaking changes

Test

from lerobot.datasets.lerobot_dataset import LeRobotDataset
from time import time

root = "/path/to/lerobot/libero_goal_image"
episodes = list(range(100))

start = time()

data = LeRobotDataset(
    "lerobot/libero_goal_image",
    root=root,
    episodes=episodes,
)

end = time()
print(f"loading {len(data.episodes)} eposides in {end - start:.2f} seconds.")
# Before: loading 100 eposides in 274.72 seconds.
# After: loading 100 eposides in 0.06 seconds.

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green