cuda: sliding-window attention, sigmoid top-k routing and combine, dense rows for every family, checked against float64

[jschmied]

What this changes

Stacked on #746 (the window test checks against its core/attention_ref.zig). Operator-only, no model calls them yet:

• window_attention.cu/.zig: sliding-window attention, decode rows over a ring of keys (split by key tile, merged
  in tile order) and prompt rows, grouped KV heads, window 0 = full.
• moe_glue.cu/.zig: top-k by logit + bias (ties: lower id), sigmoid weights, optionally normalised and scaled,
  a shared-expert slot; a warp form up to 512 experts, a block form beyond; the weighted combine of the slots.
• dense_rows.cu/.zig: bf16 rows times bf16 weight rows into fp32 (routers, output heads), a row's bits at any row
  count. core/moe_ref.zig, core/dense_ref.zig: the float64 references.

Users: families with sliding-window layers or sigmoid-routed MoE on CUDA (Kolibri 1's shapes in the tests; GLM and
DeepSeek-style routing are the same top-k form).

Receipt

• Environment: TensorFold c36b536d (1.0.9) plus cuda: tree attention and DFlash block attention for every family, checked against float64 #746 and this commit; one DGX Spark (GB10, sm_121, CUDA 13). No
  checkpoint.
• tf-cuda-test window: 48/4x128 window 513 over rings of 640 and 600 (a ring end inside a tile), 8/2x64 window 100
  with 3 streams, 16/1x128 no window; prompt rows from position 0 and across the ring's wrap: worst 0.465, bits at
  any row count and chunk.
• tf-cuda-test moe-glue: 384 experts top 6 + shared, 256 top 8 normalised x2.5, 512 top 32 (warp form's widest),
  513 top 6 (block form's narrowest), 4096 top 64; picks exact, weights and sums worst 0.996.
• tf-cuda-test dense-rows: [384, 2560], [4103, 2560], [257, 7168]: worst 0.007, bits at every row count.
• Mutations: window bound one key wider (74x), sigmoid exponent x1.0625 (5,527x), a dense row's first product
  dropped (17,529x).
• sm_120: BobClawblaw ran window, dense-rows, moe-glue on an RTX 5090, all passing with mutations (Qwen3.8-27B on CUDA: the qwen27 family's engine for NVIDIA GPUs, every stream drafted #608).
• Speed in this PR: none claimed, nothing calls it.
• Tests run: zig build test and zig build test-cuda-host (27/27 steps, 350/350 tests). zig fmt --check clean.
  lean_check.py: 538 findings, as on main.
• Not run, and why: Metal (unchanged); no model on main uses it.

Until #746 lands, this branch carries its commit; the change itself is the last commit (2710ae2d).

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (No
  serving path changes.)
• No precision traded for speed. If the bits change, the description says which sums change and why.
• Prompt processing is no slower than the last release.
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.