Fix tutorial correctness, explainer interactions, and README logo

[jxudata]

The tutorial mixed a fixed single core trace with ensemble inference terminology, and its training and inference diagrams displayed inconsistent toy outputs. Flow connectors also drifted under scaling, and lesson navigation did not update stage highlights. This fixes those behaviors and replaces the README's distorted footer mark with the official Affirm logo.

Changes:

• Specify that the Iris inspector runs quantized TabICLv2 weights in a fixed single core pass with context-only standardization, original feature/class order, and temperature 1. It is not a selected view of the playground's eight-view classifier. Explain both context-label injections and cached column attention; use a shared color scale for raw and QASSMax-scaled attention logits.
• Distinguish the nanoTabICL forward architecture from the illustrative classification training step and the released TabICLv2 training recipe. Compute inference's two-class softmax from the same toy logits used for training's ten-slot cross-entropy, label both probabilities, and explicitly identify them as illustrative rather than model predictions.
• Correct SVG connector coordinates under scaling, update lesson highlights, skip superseded query inspections, and give block/head controls accessible names.
• Synchronize PFN episode text, clear SCM rows and traces when noise changes, and replace the unsupported below-300-row claim with the revised paper's few-shot evaluation reference.
• Insert the official Affirm primary logo into the README infographic at its original aspect ratio, rounded to the nearest pixel. Retain the source asset and attribution. Pixel comparison confirms that only the footer logo region changed.

Validation:

• 45 Node tests pass, including inspection/prediction agreement and attention-weight consistency. Svelte check reports zero errors and warnings; the static build passes.
• Browser checks at 375, 768, 1024, and 1440px confirm connector alignment and no page overflow. Lesson navigation, query A/B/C, attention expansion, and block/head controls work without JavaScript exceptions. Iris probabilities remain unchanged.
• The revised inference diagram renders P₀=0.007 and P₁=0.993 with class labels at all four widths. Changing SCM noise clears the existing rows and trace; PFN task text remains synchronized.
• JS syntax, local asset references, stale-copy/visible em-dash scans, and git diff checks pass. GitHub validation passes on the latest commit (55ac1f6).

The inference graph and checkpoint are unchanged. The explorer still scales a fixed 1880px layout, making text difficult to read, including at desktop widths; that usability issue remains unresolved. Full Python parity and the notebook were not rerun. Existing dependency build warnings remain. This PR has not been merged or deployed.