[TRITON]: End-to-end fused feed-forward kernel

[willzhou-amd]

Continuation of #736.

Note that the E2E fused FF suffers from heavy atomics contention when the GEMM is large, and thus typically only outperforms the two-kernel implementation (fused GEMM+act+gating, and then another GEMM) when M<128.

Adds:

• E2E fused gated GEMM (e.g FF with swiglu)
• E2E fused ungated GEMM (e.g FF with relu)
• Benchmarks.
• Tests.

Metrics are collected on a single MI350X.

Metrics

.do_bench:

RocprofV3: