Use upper Cholesky factors

[adenzler-nvidia]

Summary

• Use Warp 1.13 upper Cholesky fill mode for tile Cholesky factorization and solves.
• Store dense Cholesky factors as upper factors and update solver Hessian/qM seeding paths to write upper-triangular storage.
• Bump the minimum warp-lang dependency to 1.13.

This should be net-same for most cases, and positive for cases where the tile size is a power of 2.

Benchmarks

Clean-machine reverify against stable baseline, three_humanoids with nworld=8192 nconmax=100 njmax=192:

metric	baseline	current	delta
mean run_time, 2 runs	7.549723	7.333066	-2.87%
mean steps/s, 2 runs	1,085,892.4	1,117,605.2	+2.92%
mean solve, 2 runs	657.815	631.126	-4.06%
Nsight total kernel time	1430.310 ms	1281.372 ms	-10.41%

Tests

• uv run ruff check mujoco_warp/_src/block_cholesky.py mujoco_warp/_src/forward.py mujoco_warp/_src/io.py mujoco_warp/_src/smooth.py mujoco_warp/_src/solver.py mujoco_warp/_src/types.py mujoco_warp/_src/smooth_test.py mujoco_warp/_src/solver_test.py mujoco_warp/_src/support_test.py
• uv run pytest mujoco_warp/_src/solver_test.py::SolverTest::test_qM_fullm_upper_indices_are_row_sorted mujoco_warp/_src/smooth_test.py::SmoothTest::test_factor_m0 mujoco_warp/_src/smooth_test.py::SmoothTest::test_factor_m1 mujoco_warp/_src/smooth_test.py::SmoothTest::test_factor_solve_i0 mujoco_warp/_src/smooth_test.py::SmoothTest::test_factor_solve_i1 mujoco_warp/_src/support_test.py::SupportTest::test_block_cholesky --tb=short -q
• git diff --check upstream/main

[thowell]

since this pr upgrades warp to 1.13.0 we can remove this guard?

mujoco_warp/mujoco_warp/_src/io.py

Line 2817 in 9fca8f7

if check_version("warp>=1.13.0.dev20260325"):

[adenzler-nvidia]

I can remove the guard, are we ok with defaulting to cuBql for all cases?

[thowell]

@StafaH can you comment on the bvh constructor? thanks!