Fix stepping past domain error for NoiseGrid when dt doesn't evenly divide time span

[ChrisRackauckas-Claude]

Summary

Fixes #136

When solving a NoiseProblem with a NoiseGrid where the dt doesn't evenly divide the time span, floating point errors could cause the solver to report "Stepped past the defined domain for the NoiseGrid".

Root Cause

This occurred because:

1. After many steps with dt=0.01 over tspan [0.0, 10.0], floating point errors accumulate and W.curt ends up at ~9.9999999999 instead of exactly 10.0
2. The step_setup flag gets set to false when W.curt + dt > W.t[end]
3. The while loop continues because W.curt < prob.tspan[2] (still slightly less due to floating point)
4. accept_step! errors because step_setup is false

Fix

The fix adds two checks in the solve loop:

1. If we're essentially at the end (within floating point tolerance), exit the loop
2. If the next step would overshoot the end, take a partial final step to reach exactly the end, resetting step_setup if necessary

Test Plan

• Added test case for the exact issue from stepping past domain #136 (dt=0.01 with grid step 0.1)
• Added test case where dt doesn't evenly divide the time span (dt=0.03 over [0,1])
• Added test case with scalar noise values
• All existing tests pass (except pre-existing failure in correlated.jl which also fails on master)

cc @ChrisRackauckas

🤖 Generated with Claude Code