This PR downgrades harmless duplicate cuDNN, cuBLAS, and cuFFT factory

registration logs from INFO to VLOG(1), fully silencing them during
normal usage.

Upstream already reduced these from ERROR to INFO, but they still create
unnecessary log noise when XLA and GPU backends initialize. Since the
duplicate registration is safe and expected this change preserves
visibility only for debugging sessions.

Co-inspired by ChatGPT during a deep dive into TensorFLow's logging
system.

Fixes: tensorflow#62075

Author: anushikhov

Diff for: third_party/xla/xla/stream_executor/cuda/cuda_blas.cc

@@ -1404,7 +1404,7 @@ void initialize_cublas() {
           });
 
   if (!status.ok()) {
-    LOG(INFO) << "Unable to register cuBLAS factory: " << status.message();
+    VLOG(1) << "Unable to register cuBLAS factory: " << status.message();
   }
 }
 

Diff for: third_party/xla/xla/stream_executor/cuda/cuda_dnn.cc

@@ -8697,7 +8697,7 @@ void initialize_cudnn() {
           });
 
   if (!status.ok()) {
-    LOG(INFO) << "Unable to register cuDNN factory: " << status.message();
+    VLOG(1) << "Unable to register cuDNN factory: " << status.message();
   }
 }
 

Diff for: third_party/xla/xla/stream_executor/cuda/cuda_fft.cc

@@ -464,7 +464,7 @@ void initialize_cufft() {
             return new gpu::CUDAFft(parent);
           });
   if (!status.ok()) {
-    LOG(INFO) << "Unable to register cuFFT factory: " << status.message();
+    VLOG(1) << "Unable to register cuFFT factory: " << status.message();
   }
 }