Add critical section to BLAS gemm.

This is needed because of possible races in SpatialConvolutionMM (and others that use gemm)
if the BLAS library is not thread-safe.

In terms of performance, there's not much benefit to run two gemms in parallel, because the
BLAS libraries have their own all-occupying gemms anyways.

generic/THTensorMath.c

@@ -1346,6 +1346,7 @@ void THTensor_(addmm)(THTensor *r_, real beta, THTensor *t, real alpha, THTensor
     m2_ = THTensor_(newContiguous)(m2);
   }
 
+#pragma omp critical(blasgemm)
   /* do the operation */
   THBlas_(gemm)(transpose_m1,
                 transpose_m2,