Barrier variant + memcpy_async benchmarks + Matmul with barrier

[louisfd]

Barrier variant

• CubeCoop: the memcpy_async call is cooperative, i.e. should not depend on UNIT_POS
• CubeManual: the memcpy_async call is per unit, can be dispatched to any unit

memcpy_async benchmarks

• Many variants on how to dispatch memcpy_async. Can beat sync loading on double buffering setup on some architecture, but otherwise the async has a lot of overhead it seems.

Matmul with barrier

• Refactoring in matmul for a copy mechanism, that works with pipeline as well as barriers for async loading. There's also a sync dummy one for testing correctness if you don't have cuda, but it's not well coalesced so very slow.
• WindowCooperativeLoading: loads all the needed slices cooperatively
• Other loadings using CubeManual are coming soon.
• Async loading do not support check bounds anymore, because the way it was done was naive and slow. A better solution coming soon.

[nathanielsimard]

If it's the same as the SimpleBarrierMatmul, I would simply make it generic on the copy mechanism instead of forking the first one.

[louisfd]

I actually tried making it generic in my subsequent upcoming PR, but it turned out to be a pain. In the end I deleted the dummy one, it's not that useful.