[Core][MultiModalHasher] Hash images without converting image mode

[lgeiger]

Purpose

This PR hashes mode, palette and data of images separately which prevents the need for converting all images to RGBA. See #24925 (comment)

Test Plan

Correctness should be covered by the existing hasher tests on CI.

The performance can be measured using:

import numpy as np
from PIL import Image
from vllm.multimodal.hasher import MultiModalHasher

np.random.seed(42)
data = np.random.randint(0, 255, size=(3840, 2160, 3), dtype=np.uint8)
data = Image.fromarray(data)

%timeit MultiModalHasher.hash_kwargs(data=data)

Test Result

For a 4k PIL image this speeds up hashing by ~35%. This is not massive, but might add up in cases with lots of multimodal input.

# main
25.1 ms ± 124 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)
# This PR
16.3 ms ± 74.6 μs per loop (mean ± std. dev. of 7 runs, 100 loops each)

[gemini-code-assist]

Code Review

This pull request improves the performance of hashing PIL.Image objects by avoiding conversion to RGBA. Instead, it hashes the image's mode, data, and palette separately. This is a good optimization that, according to your tests, speeds up hashing by ~35% for 4k images. I've found one potential issue with the implementation regarding hash uniqueness for palettized images and provided a suggestion to address it.