Fix integer overflow in generic matrix-vector multiplication

[maleadt]

The inner loop formed each product in the narrow input eltype before widening, so narrow-integer inputs overflowed. Form products in the accumulator type via muladd, as already done for matrix-matrix:

LinearAlgebra.jl/src/matmul.jl

Lines 1216 to 1220 in 27b0f13

	z2 = zero(A[i, a1]*B[b1, j] + A[i, a1]*B[b1, j])
	Ctmp = convert(promote_type(eltype(C), typeof(z2)), z2)
	@simd for k in intersect(nzcols(A, i), nzrows(B, j))
	Ctmp = muladd(A[i, k], B[k, j], Ctmp)
	end