Improve Section 04 nanoTabICL visualizations

[jxudata]

Improve Section 04's visualizations of the nanoTabICL classifier forward graph used to teach TabICLv2. The training loss and update are illustrative educational steps around that graph, not a nanoTabICL pretraining implementation. The browser model implementation, checkpoint, and prediction behavior are unchanged.

Visualization improvements:

• Draw the hidden query label and predicted distribution as separate inputs to cross-entropy, making the label's role as a training target explicit.
• Show nanoTabICL's second context-only label embedding before the ICL blocks.
• Distinguish the first eleven ICL blocks, where context and query read context K/V, from nanoTabICL's final block, which returns queries only. Connect the query's own representation to the retained query output and connect that output to LN/MLP.
• Name nanoTabICL explicitly as the schematic's source and distinguish it from the separate browser classifier's eight-view preprocessing and ensemble procedure. Remove an unused hardcoded output-bar drawing branch.

Primary architecture source: official nanoTabICL model.py at 4a7f9c7648f05c3efb34555e088105405a707469. The TabICLv2 paper v2 supplies the broader teaching context. The full TabICL implementation at 0dbff3ec8fc68c123c87af77b0ea8b25cd2d23f3 was additionally checked for data-flow consistency; it is not the implementation depicted by this schematic. Its training path can compute all row outputs and then select queries, while nanoTabICL skips unused final context outputs. The tensor dimensions also match the browser manifest, but this does not make the schematic an exact trace of that checkpoint or its wrapper. Training loss and probability values remain explicitly illustrative.

Validation: inspected the rendered forward, loss, and inference canvases; all eight steps at 375/768/1024/1440px (32 states) select correctly with no page overflow or JavaScript exceptions. Website JS syntax, local asset references, visible em-dash scan, and git diff checks pass. Subsequent label wording changes pass applicable syntax and diff checks. Core runtime and model files are unchanged.