fix tsmqr with workc usage: some gemm are done with T but lower parts of the matrix blocks was not initialized to 0

do not use cublas V2 until STARPU does not manage cublas handles - we could have a function similar to starpu_cuda_get_local_stream to get local cublas handle

do not allocate a matrix if NULL pointer is given for mat in MORSE_Desc_Create_User --> let the user manage the allocation of his data

magma_ztsqrt_gpu is now defined in magma-1.6.0: change the name of our magma routine in magma_ztsqrt2_gpu