Fix linesearch skip causing solver divergence in float32 (#1162)

The linesearch early-exit check (`|derivative| < gtol`) was unreliable
in float32 because gtol (~1e-10) falls below float32 precision (~1e-7).
This caused rare solver divergence (hitting 100 iteration limit),
bottlenecking batched simulation by ~15%.

Two fixes:
- Clamp gtol to 1e-6 minimum (matching tolerance clamp in io.py)
- Require cost improvement for the early exit, preventing acceptance
  of Newton steps that overshoot past the minimum

Author: adenzler-nvidia

mujoco_warp/_src/solver.py

@@ -1031,7 +1031,7 @@ def kernel(
     snorm = wp.sqrt(ctx_search_dot_in[worldid])
     meaninertia = stat_meaninertia[worldid % stat_meaninertia.shape[0]]
     scale = meaninertia * wp.float(nv)
-    gtol = tolerance * ls_tolerance * snorm * scale
+    gtol = wp.max(tolerance * ls_tolerance * snorm * scale, 1e-6)
 
     # p0 via parallel reduction
     local_p0 = wp.vec3(0.0)
@@ -1164,8 +1164,8 @@ def kernel(
     lo_in_sum = wp.tile_reduce(wp.add, lo_in_tile)
     lo_in = _eval_pt(ctx_quad_gauss, lo_alpha_in) + lo_in_sum[0]
 
-    # check for initial convergence: if |derivative| < gtol, accept Newton step immediately
-    initial_converged = wp.abs(lo_in[1]) < gtol
+    # accept Newton step if derivative is small and cost improved
+    initial_converged = wp.abs(lo_in[1]) < gtol and lo_in[0] < p0[0]
 
     # main iterative loop - skip if already converged
     if not initial_converged: