This simplify the *IncSubtensor code at the same time.

IncSugtensor and the GPU code still need update.

Fix grad of Alloc when we unbroadcast an input.

This is needed for a test.

In some cases it was causing an error in the grad related to
broadcasting. In the test it case bad shape.

Gpuflatten skip test

Compyte2 rebase

Fix mega typo in the Gemm -> GpuGemm optimizer.