Function roi_align 

pub fn roi_align(
    features: &Tensor,
    rois: &[(f32, f32, f32, f32)],
    output_size: (usize, usize),
    sampling_ratio: usize,
) -> Result<Tensor, DetectError>

RoI Align: bilinear-interpolation version of RoI pooling (no quantisation).

• features – [H, W, C] single-image feature map.
• rois – slice of (x1, y1, x2, y2) in feature-map coordinates.
• output_size – (out_h, out_w).
• sampling_ratio – number of sampling points per bin dimension (0 means adaptive: ceil(bin_size)).

Returns a tensor of shape [num_rois, out_h, out_w, C].