fix(wgpu): close a device profile on its last compute pass

[ThierryCantin-Demers]

Closes #1421.

A profiled region gets timestamp writes only on the pass that drains its token, so stop_profile_setup resolves the end from the set the start came from and the window is the first compute pass alone. WgpuStream closes a pass every 32 dispatches, so any longer region reports a constant.

That constant becomes zero in launch_overhead, which derives its loop length from the measurement: a window that never grows drives warmup to ~670k iterations, and 23.9 µs over that count underflows Duration to zero. It prints N/A, and roofline_bounds has been costing a launch at zero on every wgpu backend.

The fix. A profile keeps one query set for its whole span. The pass that opens it writes both slots; every later pass while it is live rewrites only the end slot, so the last pass marks the end. No extra query sets, and the allocator is untouched.

Not the per-pass sets #1421 suggests: that is ~233k create_query_set calls per sample at these counts, and under Metal's 28-set budget the round-robin hands back the set a live profile started in and clobbers that start, giving zero again.

Measured, RTX 4070 Ti SUPER, profiling N launches:

launches	main	per launch	this branch	per launch
8	7.424 µs	928 ns	7.68 µs	960 ns
64	27.136 µs	424 ns	162.8 µs	2.544 µs
512	27.648 µs	54 ns	1.917 ms	3.745 µs
4096	27.392 µs	6 ns	14.366 ms	3.507 µs

Pinned within 2 % while the work grows 64×. One and eight launches are unchanged; the single-pass regime was already right. 3.5 µs per launch matches what CUDA measures on the same card. WGSL is the same shape.

cubecl-wgpu was not instantiating testgen_profiling!, which is why this survived; it is wired in here, and test_window_spans_every_pass_in_it reports a growth ratio of 0.76 on the unfixed code while both pre-existing profiling tests pass.

Green on cubecl-wgpu --features spirv (746 + 17), on lavapipe as a CI stand-in, and on CUDA against origin/main (same 17 pre-existing failures both sides, one-test delta is the new test).

Known, not fixed here

• A cached zero outlives this fix: the throughput store has no probe-behaviour version, so a machine that recorded one keeps serving it until the cache is keyed properly.
• roofline_bounds still charges an unmeasurable launch as free; that wants Bounds::launch_overhead to become optional.
• Once Metal's 28-set budget is exhausted a new profile can open in a set an older live one started in. The QuerySetAllocator doc is corrected to say so; fixing it needs the allocator to know which sets are live.
• HIP and Metal not run. Both pick up the new test; neither is touched by this change.

Validate your PR with burn.

It is important that you make sure that you don't introduce any bugs in burn.

Instructions

• Create a new branch or fork of the cubek repo
• Update the main Cargo.toml with this PR hash.
• Fix any broken tests or compilation errors in cubek.
• Submit a PR in cubek with your fixes and link it here.
• Create a new branch or fork of the burn repo
• Update the main Cargo.toml with this PR and the cubek PR hash.
• Fix any broken tests or compilation errors in burn.
• Submit a PR in burn with your fixes and link it here.