Vector / chunked / batch mode

[gdalle]

This one is rather easy in principle: just wrap the JVP/VJP/HVP seed in a Batch struct that is essentially a named tuple, and dispatch on that

[gdalle]

Partially solved by #320, still needs:

• Batched HVPs based on batched pushforward/pullback (done in Batched HVP #330 except for reverse-over-forward)
• Batched pushforward with ForwardDiff (done in Batched ForwardDiff pushforward #328)
• Batched pushforward and pullback with Enzyme
• Tests on efficiency (type stability, allocations)
• Independent tests of the primitive operators with batched scenarios (done in Test batched operators #331)

[gdalle]

It would be nice to avoid the distinction between operator and operator_batched, but that would require wrapping inputs into a Single struct or so

[gdalle]

Further improved by #412, which makes batch mode the default that extensions need to implement

[gdalle]

Batch mode for Enzyme is nearly complete since #471, it only needs EnzymeAD/Enzyme.jl#1883 to work with out-of-place functions