gbmv! prone to NaNs in y

[dlfivefifty]

This behaviour is inconsistent (and prevents use of similar):

alpha = T(2.0)
beta = T(0.0)
hA = rand(T, 4, 32)
hB = rand(T, 32)
m = 33
hC = rand(T, m)
A = CLArray(hA)
B = CLArray(hB)
C = CLArray(hC)
isnan(sum(BLAS.gbmv!('N', m, 1, 2, alpha, A, B, beta, C))) # false
C[1] = NaN
isnan(sum(BLAS.gbmv!('N', m, 1, 2, alpha, A, B, beta, C)) ) # true
hC[1] = NaN
isnan(sum(BLAS.gbmv!('N', m, 1, 2, alpha, hA, hB, beta, hC))) # false