Inefficient implementation of LBFGS multiply

[MaxenceGollier]

LinearOperators.jl/src/lbfgs.jl

Lines 195 to 197 in 4575ca5

	for j ∈ eachindex(q)
	q[j] += bx * data.b[k][j] - ax * data.a[k][j]
	end

This is highly inefficient. There are too many memory accesses being performed. We should replace with

@. q += bx .* data.b[k] - ax .* data.a[k]

In the same function there is also

LinearOperators.jl/src/lbfgs.jl

Line 190 in 4575ca5

for i = 1:(data.mem)

which should be improved with the @inbounds macro.

@dpo I am already getting huge improvements just with this (regarding the discussion we had this morning on Zulip).

[dpo]

Thanks. That code was probably written before the dot notation existed.