Add New Training Method

Add new training method:

1. Add loss function to mip/losses.py
2. Add corresponding sampler to mip/samplers.py
3. Register both in get_loss_fn() and get_sampler()
4. Test with unit tests

Samplers

Samplers in mip/samplers.py implement inference-time sampling strategies that correspond to each loss type:

def sampler(
    config: OptimizationConfig,
    flow_map: FlowMap,
    encoder: BaseEncoder,
    act_0: torch.Tensor,  # Initial noise
    obs: torch.Tensor,    # Observations
) -> torch.Tensor:
    """
    Returns:
        action: Predicted action trajectory
    """

Losses

All loss functions in mip/losses.py follow a consistent signature:

def loss_fn(
    config: OptimizationConfig,
    flow_map: FlowMap,
    encoder: BaseEncoder,
    interp: Interpolant,
    act: torch.Tensor,
    obs: torch.Tensor,
    delta_t: torch.Tensor,
) -> tuple[torch.Tensor, dict]:
    """
    Returns:
        loss: Scalar loss tensor
        info: Dict with additional metrics for logging
    """