mlx-vlm Remove scipy dependency

Its a heavy dependency which requires external libraries and tooling installed. We are only using it for the scipy.ndimage.zoom method which is provided both by Pillow and OpenCV which are already a depepndency for this project.

Of the top of my head these two implementations can replace zoom method, note that the opencv one will be the best performant:

from PIL import Image
import numpy as np

def pillow_zoom(input_array, zoom_factor):
    """Resize using Pillow with similar functionality to scipy.ndimage.zoom"""
    if input_array.ndim == 2:
        # Handle 2D arrays (single-channel images)
        img = Image.fromarray(input_array)
        new_size = tuple(int(s * zoom_factor) for s in input_array.shape)
        # Swap dimensions for PIL (width, height)
        resized = img.resize((new_size[1], new_size[0]), Image.Resampling.BICUBIC)
        return np.array(resized)
    elif input_array.ndim == 3:
        # Handle multi-channel images
        results = []
        for i in range(input_array.shape[0]):
            results.append(pillow_zoom(input_array[i], zoom_factor))
        return np.stack(results)
    else:
        raise ValueError("Unsupported array dimension")

or

import cv2
import numpy as np

def opencv_zoom(input_array, zoom_factor):
    """Resize using OpenCV with similar functionality to scipy.ndimage.zoom"""
    if input_array.ndim == 2:
        # Single channel 2D array
        new_shape = tuple(int(s * zoom_factor) for s in input_array.shape)
        # Note: cv2.resize takes (width, height) which is opposite of numpy's (height, width)
        return cv2.resize(input_array, (new_shape[1], new_shape[0]), interpolation=cv2.INTER_CUBIC)
    elif input_array.ndim == 3:
        # Multi-channel or 3D array
        results = []
        for i in range(input_array.shape[0]):
            results.append(opencv_zoom(input_array[i], zoom_factor))
        return np.stack(results)
    else:
        raise ValueError("Unsupported array dimension")