tests: re-measured robot tolerance overrides (36 -> 15)

[plancherb1]

Re-measured every per-robot override in tests/tolerances.py against Pinocchio in fp64 (full suite with --runslow, logging each comparison against the default bucket).

• Deleted 21 overrides whose default bucket already passes with >= 2x margin: all 9 g1 buckets, 6 of the 8 h1_2 buckets (e.g. aba atol 6e-2, minv atol 1.0 "cond(M)~5e6"), fetch aba/inverse_dynamics, fr3 centroidal, iiwa14 pose_hessian, rizon4 inverse_dynamics, gen3 f_ext_gradient_so. After the floating-mimic fixes RBDReference matches Pinocchio on h1_2 to ~1e-15.
• Tightened h1_2 inverse_dynamics (1e-3/1e-2 -> 3e-4/3e-6; only the fp64 RK4 integrator cells need slack) and h1_2 second_order_fdsva (atol 0.1 -> 3e-3; one near-zero entry of 91,125 misses by 1e-3).
• Kept the documented gen3 continuous-joint buckets and baxter/rizon4 aba.

Verified: full suite with --runslow passes with the new table. GRiD's CUDA tests that read these buckets were checked too (the g1 f_ext test was the only one leaning on a deleted large atol; it passes).

🤖 Generated with Claude Code