Add GraND and GREAT batch selection strategies

[jainarchita]

Summary

• Implement GraND (Gradient Normed Distance) batch selection strategy that selects training samples based on gradient norm magnitudes, combining exploration (random sampling) with exploitation (high gradient norm samples)
• Implement GREAT (GREedy Approximation Taylor) batch selection strategy that greedily selects samples with maximum orthogonal gradient contribution to maximize expected loss reduction
• Unify the batch sampler interface so all strategies (including existing ones) use a consistent **kwargs pattern for passing model, loss function, device, and loss history

Changes

• New files: trainer/batching/vision_batching/grand.py, trainer/batching/vision_batching/great.py
• Modified: trainer/pipelines/vision/vision.py : unified batch sampler call interface, added GPU support and device info logging
• Modified: trainer/constants.py : increased default runs to 5, added automatic CUDA device detection
• Modified: trainer/constants_batch_strategy.py : registered GraND and GREAT strategies
• Both strategies use a candidate pool approach (similar to MILO/CORESET) for computational efficiency, with a reduced pool size (500) for GREAT given its higher per-sample cost