tests: fix idefics3 SmolVLM processor size in multimodal tokens test
Remove unnecessary setUpClass caching; instead set size=1024 (2x2=5 tiles)
directly in the test body to reduce image processing cost vs default 2048
(4x4=17 tiles), cutting tile work ~3x without caching the processor.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>