tp: apply parabolic half-accel only during an actual blend

[grandixximo]

Fixes #4220.

What

The trajectory planner halves acceleration on segments that never blend (single MDI move, first segment from rest, last segment decelerating to a stop). See #4220 for the full analysis and evidence.

This gates the half reduction on an actual parabolic blend overlap instead of a term-condition flag.

Change

• src/emc/tp/tc.c: keep tcGetOverallMaxAccel / tcGetTangentialMaxAccel reduced (used for corner-speed sizing, unchanged). Add tcGetCycleMaxAccel(tc, in_overlap) that applies the half only when the segment overlaps a neighbor in an active blend, otherwise returns the full path acceleration.
• src/emc/tp/tp.c: thread an in_overlap flag through tpUpdateCycle and the trapezoidal / ramp / s-curve per-cycle accel functions. Overlap is true for the secondary segment of an active blend and for the primary while blending_next is latched; false for lone / first-from-rest / last-to-stop portions and tangent hand-offs.
• tests/motion/g0: updated to expect full acceleration on a lone rapid (it previously asserted exactly half).

Result

• Lone MDI / one-line program: full acceleration.
• First segment: full from rest, half only once it blends into the next.
• Last segment: half during the entry blend, full for the deceleration to a stop.
• Real blend overlap: both segments still half, so per-axis sums stay within limits and corner speeds are unchanged.

The full path acceleration from canon already respects every joint limit for the segment direction, so a single moving segment at full accel violates nothing.

Testing

runtests tests/abort tests/motion tests/lathe: all pass with the g0 update. Blend, stop, abort, and lathe behavior is unchanged.