perf: Use OpenCV over PIL for PNG encoding in ImageRef.from_pil

[maxdswain]

Overview
The ImageRef.from_pil class method is used widely in docling's codebase. It is often used several times per page when parsing documents in the docling_parse.pdf_parser.PdfDocument._to_bitmap_resources_from_decoder method. From my profiling, I found that it took up ~45% of processing time when doing a DocumentConverter conversion with all AI models disabled. This led me looking into how it's performance can be improved.

The function uses pillow to encode the image to a png, which is notoriously slow. So I swapped it out with opencv, improving the performance of this function by ~55% for this simple test case:

import timeit

setup = """
from PIL import Image as PILImage
from docling_core.types.doc import ImageRef
"""

stmt = """
fig_image = PILImage.new(mode="RGBA", size=(200, 400), color=(0, 0, 0))
image = ImageRef.from_pil(image=fig_image, dpi=72)
"""

result = timeit.timeit(stmt=stmt, setup=setup, number=1000)
print(f"1000 runs: {result:.4f}s")
print(f"Per call:  {result:.4f}ms")
# Before: 1000 runs: 3.9219s, Per call:  3.9219ms
# After: 1000 runs: 1.7429s, Per call:  1.7429ms

When using these changes in the main docling repo, it reduced by conversion time from 14.2 to 9.21 (~35%) when disabling all AI models.

One caveat is that I did add an extra dependency opencv-python-headless, however this is already a dependency in the main docling repo's uv.lock.