[AMD] Default to stream-pipeline-v2

[sjw36]

No description provided.

[binarman]

I've looked into LDS overflow issue in this PR: pipeline allocates memory for gemm operands like this:

%47 = triton_gpu.local_alloc  : () -> !tt.memdesc<2x128x128xf8E5M2, #shared, #triton_gpu.shared_memory, mutable>
%48 = triton_gpu.local_alloc  : () -> !tt.memdesc<2x128x256xf8E5M2, #shared1, #triton_gpu.shared_memory, mutable>

First allocation takes 2x128x128=32 kBytes
Second allocation takes 2x128x256=64 kBytes

This happens because num_pipeline_stages kernel argument equal 3 by default.

I think we can simply reduce this value to 2 for AMD, LDS is sufficient in this case.

[antiagainst]

LGTM! But blocking for now before figuring out some final perf regressions