More intuitive behavior of exists() for frame fields

[brimoor]

Updates the logic of the exists() stage to be more intuitive for frame fields:

import fiftyone as fo

sample1 = fo.Sample(filepath="video1.mp4", index=1)
sample1.frames[1] = fo.Frame()

sample2 = fo.Sample(filepath="video2.mp4", foo="bar", index=2)
sample2.frames[1] = fo.Frame(foo="bar")

sample3 = fo.Sample(filepath="video3.mp4", index=3)

dataset = fo.Dataset()
dataset.add_samples([sample1, sample2, sample3])

# Only contains samples with a `foo` attribute
view = dataset.exists("foo")
assert view.values("index") == [2]

# Only contains samples without a `foo` attribute
view = dataset.exists("foo", bool=False)
assert view.values("index") == [1, 3]

# Only contains samples with >= 1 frame
view = dataset.exists("frames")
assert view.values("index") == [1, 2]

# Only contains samples with no frames
view = dataset.exists("frames", bool=False)
assert view.values("index") == [3]

# Only contains samples with a `foo` attribute on >= 1 frame
view = dataset.exists("frames.foo")
assert view.values("index") == [2]

# Only contains samples with no `foo` attribute on any frames
view = dataset.exists("frames.foo", bool=False)
assert view.values("index") == [1, 3]

[benjaminpkane]

LGTM