Remove redundant bounds check in mul! with sparse matrices

[haampie]

There's performance to be gained in sparse mat-vec products for very sparse matrices by moving the @inbounds one block up. We can probably do this cause the dimensions of B and C have already been checked.

Some results:

function bench(n = 1_000, k = 5)
    # Tridiagonal matrix
    A = spdiagm(-1 => fill(-1.0, n-1), 0 => fill(2.0, n), 1 => fill(-1.1, n-1))
    B = ones(n, k)
    C1 = similar(B)
    C2 = similar(B)

    mul!(C1, A, B, 1.0, 0.0)
    my_mul!(C2, A, B, 1.0, 0.0)

    @assert C1 == C2

    mul!(C1, A', B, 1.0, 0.0)
    my_mul!(C2, A', B, 1.0, 0.0)

    @assert C1 == C2

    A1 = @benchmark mul!($C1, $A, $B, 1.0, 0.0)
    A2 = @benchmark my_mul!($C1, $A, $B, 1.0, 0.0)

    Aᵀ1 = @benchmark mul!($C1, $(A'), $B, 1.0, 0.0)
    Aᵀ2 = @benchmark my_mul!($C1, $(A'), $B, 1.0, 0.0)

    A1, A2, Aᵀ1, Aᵀ2
end

n	k	old A * x	new A * x	old A' * x	new A' * x
1_000	5	22μs	17μs	21μs	18μs
10_000	5	222μs	175μs	207μs	182μs
100_000	5	2395μs	1933μs	2385μs	2078μs