Allow floating point constants in ImageMath operations

[vera-rubin]

ImageMath.lambda_eval raises TypeError: color must be int or single-element tuple when we add 0.5 to an L image. The documentation says this should produce an F image, so a pixel of 10 should become 10.5.

from PIL import Image, ImageMath

a = Image.new("L", (1, 1), 10)
out = ImageMath.lambda_eval(lambda args: args["a"] + 0.5, a=a)
print(out.mode, out.getpixel((0, 0)))

Changes proposed in this pull request:

We now create integer scalar images only for integer constants, allowing floats to use the existing promotion to mode F.

We checked the regression, both existing ImageMath suites, and the Pillow selftest.

Tested with Pillow 13.0.0.dev0 (bc7e6f5), Python 3.12.3.

— ChipLab