Feat/matmul/simple barrier

[louisfd]

• Add cooperative barrier option, in which all the cube must perform all memcpy_async calls.
• Add 5 loaders working with the strided layout. Some of them are experimental, will probably be deleted at some point
• Make loaders generic over the copy mechanism (pipeline vs barrier)
• Since i'm experimenting, i fork simple_matmul a bit. This will be cleaned up once it's more clear what we want to keep
• Fix memcpy to use references

[louisfd]

I'll make another PR containing this one