perf(cuda): make GPU kernel launches asynchronous (drop per-launch device sync)

[isPANN]

Problem

Every tropical-GEMM kernel launch helper in tropical-gemm-cuda issued a blocking stream.synchronize() immediately after enqueuing the kernel:

unsafe { builder.launch(cfg)?; }
stream.synchronize()?;   // <-- full host<->device round-trip, every launch

For a single large matmul this is invisible. But for a tropical tensor-network contraction made of thousands of tiny nodes (e.g. max-plus weighted-MIS), it serializes the entire pipeline — the host blocks on each ~few-µs kernel while the GPU sits idle between launches.

Evidence

An nsys CUDA trace of a 9431-node max-plus contraction on an A40:

• 42,442 cuStreamSynchronize calls — exactly equal to the 42,442 GEMM kernel instances (one blocking sync per launch).
• Actual GPU kernel time was only ~36% of the wall; the rest was the host stalled in these per-launch syncs.
• A network with the same total arithmetic packed into a few large nodes was GPU-bound and fast — confirming the gap is per-launch sync overhead, not kernel throughput.

Fix

Remove the per-launch stream.synchronize() from every kernel-launch helper (forward GEMM, argmax, external/DLPack, backward, and K-packed AndOr), making launches asynchronous on the context's stream.

This is correctness-preserving:

• All operand uploads, kernels, and device-to-device copies run on the context's single stream (CudaContext::stream() = the device default stream), so stream ordering guarantees each launch observes its predecessors' results.
• Every host read goes through GpuMatrix::to_host, which synchronizes — those two barriers are kept.
• Callers that want an explicit barrier (e.g. before timing) can call ctx.stream().synchronize().

A module-level doc comment in kernels.rs documents this asynchronous-launch contract.

Notes

• Base is feat/cuda-andor-gpu (the branch downstream omeinsum-rs / miso pin); retarget to main if preferred.
• No API changes; +34 / -10, comments + doc only beyond the removed syncs.
• CUDA build/tests validated downstream on A40 (can't compile the cuda feature on non-CUDA CI hosts).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.80%. Comparing base (82d2edb) to head (7b713cc).

Additional details and impacted files

@@                 Coverage Diff                  @@
##           feat/cuda-andor-gpu      #59   +/-   ##
====================================================
  Coverage                94.80%   94.80%           
====================================================
  Files                       20       20           
  Lines                      944      944           
====================================================
  Hits                       895      895           
  Misses                      49       49           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[isPANN]

Superseded by the main-targeted PR above: the per-launch sync removal belongs on main (the launch helpers are shared; main already carries the device-resident API that downstream omeinsum/miso use). The K-packed AndOr launchers on this branch get the identical one-line change when #44 lands on main. Closing in favor of the main PR.