materials: build the deviatoric projector once, not per call (#77)

[petlenz]

Closes #77.

Both generic yield-function policies constructed the rank-4 IIdev inside flow_normal_stress_derivative(), which runs once per plastic step:

const tensor4 IIdev{plasticity_detail::make_IIdev<T, Dim>()};

It depends on nothing but T and Dim.

Measured

-O2 -DNDEBUG, paired/interleaved timing (from the performance review on #77):

	fresh	cached
make_IIdev<double,3>()	149 ns	2.6 ns

plastic step	before	after
drucker_prager_plasticity	700 ns	548 ns	-22%
j2_rk_plasticity (implicit_euler)	971 ns	759 ns	-22%
j2_rk_plasticity (sdirk3)	1119 ns	953 ns	-15%

j2_plasticity had always cached it in a member, which is most of why its plastic step was 3-6x cheaper than the two materials that go through the policies. This brings them to the same standard.

Why a function-local static

The policies are value types that get copied around, so a member would be rebuilt per copy. This way one instance is shared by every material at that (T, Dim). Initialization is thread safe, which matters because the UMAT layer runs one context per thread; after the first call only the guard check remains, and that is inside the 2.6 ns above.

Correctness

Stress and tangent are unchanged -- the existing J2, Drucker-Prager and RK value assertions all still pass. 311/311.