no samples shown for language benchmarks

for vision models, we show sample images from the benchmark. Can we do the same with language stimuli?
At the very least, we should get rid of the `sample` placeholders (see e.g. http://www.brain-score.org/model/language/1651):

Open the linked language benchmark page and compare it with the existing vision benchmark sample display. Trace the corresponding benchmark-page rendering flow, then make language stimuli appear as samples or remove the `sample` placeholders. Done means language benchmark pages no longer show empty placeholders and display samples when available.

Written by the indexing model from the issue text.