VecMat fusion autotune bug

[laggui]

Running wgpu tests fails due to vecmat fusion autotune:

cargo test  --lib --bins --examples -p burn-wgpu --color=always --release

  thread 'tests::cube_fusion::autodiff::f32_ty::ad_transpose::tests::should_diff_swap_dims' panicked at /home/agent/.cargo/git/checkouts/cubecl-058c47895211d464/5a4ad7f/crates/cubecl-runtime/src/tune/local.rs:155:26:
  Should run when selected by autotune.: Unknown("RunnerError(LaunchError(Unable to launch matmul because the config is invalid: \"Lhs and Rhs must have same line size, got lhs=1 and rhs=2\"\n))")

When running the failing test should_diff_swap_dims only, from a clean cache, it passes (but if you run it using the current cache it will fail, as expected):

cargo test  --lib --bins --examples -p burn-wgpu --color=always --release should_diff_swap_dims

And if we remove (comment out) the SimpleVecMat and DoubleVecMat tunables when running all the tests, there are no issues.

burn/crates/burn-cubecl-fusion/src/matmul/tune.rs

Lines 84 to 89 in ff5f34e

	TunableSet::new(create_key::<R>, input_gen::<R>)
	.with(Tunable::new(tune_fallback::<R, BT>)) // First one should always work.
	.with(Tunable::new(tune_fused::<R, BT, SimpleUnit>).group(&unit, |_| PRIORITY_MAX))
	.with(Tunable::new(tune_fused::<R, BT, SimpleVecMat>).group(&unit, |_| PRIORITY_MAX))
	.with(Tunable::new(tune_fused::<R, BT, DoubleVecMat>).group(&unit, |_| PRIORITY_MAX))
	.with(

It looks like the vecmat are selected by a previous test during autotune, so when running should_diff_swap_dims it is selected as the fastest but fails to execute at runtime due to line size mismatch.

[wingertge]

It seems to be introduced by this PR, or at least the error doesn't happen before. Since the PR specifically mentions changing the swapped dimension logic it seems related.
#3627