[Feature] swimlane_converter: render SPMD deps at block granularity, not a per-instance crossbar

[lyfne123]

Summary

The runtime resolves SPMD inter-block dependencies at per-SPMD-block (scope) granularity, but swimlane_converter expands each block-level edge into a dense per-instance crossbar of flow arrows. Represent the dependency at block granularity (or clearly annotate it) so the trace doesn't imply per-instance dependency resolution.

Motivation / Use Case

• For a whole kernel (DSv4 decode_sparse_attn), deps.json captures the dependency graph as 9 task nodes / 23 tensor-mediated edges (source: tensormap, region overlap: covered/other) — i.e. one node per SPMD block.
• The same run's merged_swimlane_*.json contains 613,820 flow events (cat: flow, name: dependency) — every block-level edge is blown up into a per-instance crossbar (e.g. 128 producer instances × 64 consumer instances).
• Empirically the runtime does a per-scope barrier: dependent stages run sequentially with only a small dispatch gap and do not overlap at the instance level, while a dependency-free input-only pre-pass correctly floats ahead and overlaps. So the per-instance arrows do not correspond to real runtime edges.
• This misleads performance analysis — a reader looking at the swimlane concludes the runtime resolves a full per-instance dependency between consecutive SPMD stages, and reasons about dependency-resolution cost / overlap potential incorrectly. It also bloats the merged trace from ~2 MB to ~110 MB.

Proposed API / Behavior

Any subset:

1. Collapse the per-instance flow arrows to one representative edge per (producer-block, consumer-block) pair, annotated with the fan degree (e.g. 128→64).
2. Add a block-level dependency view / toggle alongside the per-instance one.
3. At minimum, document (tool help + output) that the flow arrows are a conservative visualization expansion of block-level tensormap edges, not runtime per-instance edges.

Reducing the arrow count would also shrink the merged trace dramatically.

Additional Context

• Repro: run a kernel with --enable-l2-swimlane --enable-dep-gen (e.g. pypto-lib models/deepseek/v4/decode_sparse_attn.py -p a2a3), then compare deps.json (9 nodes / 23 edges) against merged_swimlane_*.json (613,820 flow events) in the same dfx_outputs/.
• runtime (simpler) commit: 48980572.
• Related: simpler#1001 (text-based dep_gen rendering) — adjacent tooling, but a different concern (output format of the dep_gen graph).