Replace the flex metric factorization with a block preconditioner

[smallquail]

Every step, the flex block of the effective metric M + K was factorized by sparse Cholesky, because K depends on the configuration. This replaces that factorization with the metric's per-vertex 3x3 diagonal blocks, prefactored.

Where the time goes

model/flex/bag.xml (added here, 2868 dof), on upstream:

	ms/step	share
build + factorize the metric	9.64	50%
constraint solve, 39 CG iterations	9.16	48%
everything else	0.43	2%
total	19.23

The factorization goes away entirely. The blocks are O(n) to build and to apply, but weaker: CG runs about twice the iterations, and qacc_smooth becomes an iteration rather than a direct solve. Net 19.23 -> 8.29 ms/step.

Preconditioner, by case

The metric has two consumers, the CG constraint solver (Mgrad = Mtilde \ grad) and the qacc_smooth solve, and both apply the same operator.

metric state	upstream	this PR
inactive: no flex elasticity, or an explicit integrator	M^-1	unchanged
bending only, nefmK == 0	M^-1, plus the exact constant bending factor from mj_setConst on the dofs it covers	unchanged
per-step stiffness, exactly factorizable	M^-1, plus a per-step sparse Cholesky of the flex block	M^-1, plus the per-vertex 3x3 blocks
per-step stiffness, M couples across the flex block	an inner PCG of up to 50 iterations over the row above, run once per outer CG iteration	M^-1, plus the per-vertex 3x3 blocks

The constant bending factor is factored at model compile time and costs nothing per step, so it stays.

Models affected

model	upstream	this PR
bag.xml, per-step stretch	19.23 ms/step, 39 CG	8.29 ms/step, 86 CG
poncho.xml, bending	2.96 ms/step, 16.8 CG	2.96 ms/step, 14.6 CG
poncho_edgeequality.xml, bending	2.11 ms/step, 9.8 CG	2.11 ms/step, 9.1 CG
flag.xml, trampoline.xml, metric inactive	unchanged path	unchanged path

Only models carrying per-step stretch stiffness change in wall-clock. Bending-only models hold their timing and take slightly fewer CG iterations, because the preconditioner is now symmetric: it applies M^-1 and the covered blocks to disjoint sets of dofs, where before the two overlapped and the operator was not symmetric, which PCG requires.

Also in this change

• mjd_effSolve is renamed mjd_effPrec, since it is a preconditioner rather than a solve, and the name mjd_effSolve now belongs to the function that solves (M + K)x = b. The qacc_smooth call site in mj_fwdAcceleration is textually unchanged but now reaches the iterative solve.
• That solve converges on opt.tolerance rather than a hardcoded 1e-10, and raises mjWARN_INERTIA on reaching its iteration cap rather than returning an under-converged qacc_smooth.
• Covered dofs are located by walking the covered rows of the stiffness matrix. They are not necessarily 3-aligned from dof 0: any joint declared before a flexcomp shifts them.
• model/flex/bag.xml is added because no shipped model carried per-step stretch stiffness. The ponchos are bending-only, and trampoline.xml uses an explicit integrator, so the metric never activates there.

In mjUSESINGLE builds the default opt.tolerance of 1e-8 is below float epsilon, so the solve cannot reach it and would warn on every flex model. A floor on the tolerance would address that; it is not included here.