Cache inner LinearCache in DualLinearCache in direct dual path

[hersle]

A suggested fix to #1064. This is largely generated with Claude. I admittedly don't understand the internals well enough to know whether this is the right fix to make. Any help is appreciated :)

It fixes the original example in #1064 (comment), which now gives:

@time ForwardDiff.derivative(usumend_klu, 1.0) # 0.000808 seconds (1.98 k allocations: 300.891 KiB, 0.09% compilation time: 100% of which was recompilation)
@time ForwardDiff.derivative(usumend_pureklu, 1.0) # 0.000636 seconds (1.93 k allocations: 315.426 KiB, 0.05% compilation time: 100% of which was recompilation)

Now PureKLU is faster than KLU (as it should be?).

[hersle]

This is my best attempt for now. I think the problem was that A was being refactored at every single solve! even when A is unchanged between the solves. Is this PR a reasonable way to fix that, or is there a better way that achieves the same?

The changes here make PureKLU keep using the direct AD path and makes it faster than KLU. #1066 instead moves PureKLU to the split AD path, which makes it only as fast as KLU.