Closes #893

I made a pass though the stress routines:

• Only compute 6 instead of 9 components
• Avoid generic FFTs when building the density
• Be more careful when caching projectors (caching should in general ignore the dual tag and just care about the value of the G vector norm)
• Use explicit derivative formula in custom rule to avoid duals in integration routines, such that loop vectorisation can play its tricks
• Introduce energy-only function, which avoids building the Hamiltonian object (and allocating its cache for matrix-vector products)

Saves a factor of 3 in runtime for the stresses on of my tests and drastically reduces allocations (about a 5th).