fix(cuda): drop temporary fast_fp16_available GCN4 diagnostic

719476c bundled a legitimate wave64 shuffle-width fix with an
unrelated, self-labeled temporary diagnostic that force-disabled the
fast fp16 path on GCN4/Polaris to test an accumulated-precision-margin
theory. Flagged by the SP2+SP2.5 final whole-branch review as leaked
WIP with no place in the merge. Revert fast_fp16_available to its
pre-diagnostic behavior; the shuffle-width fix itself is untouched.

Author: kmbandy

ggml/src/ggml-cuda/common.cuh

@@ -302,9 +302,7 @@ static bool fp16_available(const int cc) {
 }
 
 static bool fast_fp16_available(const int cc) {
-    // DIAG(temp): gfx803/Polaris (GCN4) fp16 ALU is not truly full-rate/clean; force the
-    // fp32-accumulation kernel paths there to test the accumulated-precision-margin theory.
-    return (GGML_CUDA_CC_IS_AMD(cc) && !(GGML_CUDA_CC_IS_GCN(cc) && cc <= GGML_CUDA_CC_GCN4)) ||
+    return GGML_CUDA_CC_IS_AMD(cc) ||
         (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc));
 }