[llvm] [AMDGPU] Move tryEffectiveStall to GCNSchedStrategy class, optionally use in tryPendingCandidate (PR #191924)

[Kevin Choi via llvm-commits]

choikwa wrote:

The test aims to show that DS_READ Defs are immediately used by MFMAs. In order to create longer live range for latency hiding, we need to 1) enable using stalls in tryPendingCandidate, 2) determine if DS_READs are taking longer cycles via heuristics or user-request, and 3) estimate and use higher latency value for DS_READs. Ideal case should show DS_READs scheduled in advance (+1 MFMAs) for most latency hiding.

The motivation came from scheduling for matmul fp8 triton benchmark.

https://github.com/llvm/llvm-project/pull/191924