Add GroupTransform module for differentiable group-wise time shift/scale/warp

[denix56]

Motivation

• Add a learnable, differentiable group-wise temporal transform to enable per-group time shifts, amplitude scaling/bias, and monotone time warping for augmentations or learned transformations.
• Initialize parameters near identity and allow configuration via group_cfg to make transforms weak by default and easy to tune.

Description

• Add timesead_ext/models/transforms/group.py implementing GroupTransform that supports differentiable time shifting via grid_sample, amplitude scale/bias (softplus-based scale and tanh-bounded bias), and monotone warping using cumulative softplus knots resampled to length T with F.interpolate.
• Parameters are learnable and initialized near identity using a _softplus_inverse helper for scale/knots and zeros for shift/bias, and they are read from group_cfg keys: k_group, enable_shift, enable_scale_bias, enable_warp, warp_knots, max_shift_frac, scale_range, bias_range, and warp_strength.
• Implements internal helpers for building sampling grids (_make_base_grid, _get_shift_grid, _get_warp_grid), a 1D grid_sample wrapper _grid_sample_1d, and broadcasting utilities for group/ batch expansion.
• Expose the new transforms package in timesead_ext/models/__init__.py and add timesead_ext/models/transforms/__init__.py exporting GroupTransform.

Testing

• No automated tests were run as part of this change.

---

Codex Task