stop clipping rejected steps to t1

[jpbrodrick89]

• Fixes Infinite final-step rejection loop due to conflicting endpoint clipping and adaptative RK step rejection #756

_clip_to_end previously overrode the PID controller on a rejected step with a fixed midpoint that is also rejected; on rejection tprev doesn't advance, so the midpoint is constant, and the same too-large step gets retried forever.

The midpoint override's only justification was floating-point stability of dense interpolation against the old fixed 1e-6 tolerance. #660 replaced that with 100 * ULP(t1), which is directly related to the float resolution boundary. This PR just drops that branch and trusts the controller's shrunken proposal on rejection. The #85 fix (only snap-to-t1 on accepted steps) is preserved.

Regression coverage:

• _clip_to_end unit test pinning the three branches directly.
• End-to-end test: harmonic oscillator at t1=600 in float32 with rtol=1e-4 deterministically hangs at the clip floor (final_t ≈ 599.9937) on main and finishes in ~30 steps with this patch.

[patrick-kidger]

LGTM! Merged 🎉