Read ps3_scales, not s3_scales, from the PS3 vision encoder

[appleweiping]

Fixes #282.

PS3VisionEncoder in ps3-torch defines ps3_scales; there is no s3_scales attribute on it, so every access of the misspelled name raises AttributeError as soon as a PS3 model is built — config.ps3=True fails before training or inference can start.

#282 reports the occurrence in multimodal_encoder/builder.py. The same misspelling occurs four more times in llava_arch.py's encode_images_ps3(), on the path reached when ps3_dynamic_aspect_ratio is enabled, so this PR fixes all five.

Two tests are included that run without torch or ps3-torch installed: one builds a PS3 vision tower against a stubbed encoder and asserts mm_scale_num is derived from ps3_scales; the other scans the source for the misspelling, which keeps the four dynamic-aspect-ratio occurrences covered without needing a GPU. The scan matches .s3_scales specifically and leaves the unrelated s2_scales used by the dynamic-S2 path alone. Both tests fail against main and pass with this change.

I have not been able to run a full PS3 model end-to-end (no GPU or ps3-torch in my environment); verification is against the ps3-torch source plus the tests above.

🤖 Generated with Claude Code

[appleweiping]

Validated the latest head b408d23 locally: pytest -q tests/test_multimodal_encoder_builder.py -> 2 passed. The tests cover the ps3_scales attribute and source-level protection against the old s3_scales typo without requiring torch/PS3/GPU. The five typo fixes are ready for review.