[grandixximo]

Fixes #4805.

The tests from #4730 waited on nothing but motion and polling. The sims moved at 90 deg/s on the rotaries and 120 mm/s on the slides, often under TCP, where the slides swing around a pivot to keep the tip still, and every step waited out a 10 ms task cycle, the 10 ms poll of wait_complete() and two 50 ms status polls. twp-table did not stall: it finishes after about 5.6 minutes here.

test	before	after
twp-native	70 s	7.1 s
twp-native-ptp (split from twp-native)		7.2 s
twp-table	338 s	9.0 s
twp-table-one-sided (split from twp-table)		4.1 s
ptp-machine-frame	26 s	5.5 s
ptp-robot	22 s	4.2 s
kins-switch	16 s	8.4 s

About 3.6 s of each is LinuxCNC starting and stopping. All of them pass with six copies of each running beside four busy loops, and pinned to one CPU.

• The sims move fast. Most checks look only at where a move ends. twp-native's G30.5 check, which reads the path from a sampler once a servo cycle, now derives how far short of the corner a sample can fall from the joint's own acceleration, instead of a number written for 700 mm/s^2.
• The tests wait on what wait_complete() waits for, polled every 2 ms, and task runs a 1 ms cycle, except in kins-switch (below).
• The faster polls showed three races in the tests that the slow polls had hidden, now fixed:
  • in position with an empty queue is also true before task has sent a move on, so the tests also wait for task's interpreter to be idle;
  • two polls can read the same status copy, so a move counts as ended only on a copy written two task cycles after the command, and still joints only across two copies with motion's heartbeat apart;
  • twp-native's sampler log is block buffered, so it now waits for the log to catch up before a move as well as after, and reads only the new lines.
• twp-table and twp-native are split in two, so each half starts LinuxCNC once and the halves run in parallel. twp-native-ptp shares twp-native's INI and HAL files through symlinks and its helpers through twp-native/native.py.
• twp-table starts each case with one joint-space G69 G53.5 G0 move, checks "no P is the nearer pose" against the P1 and P2 poses it already has from the same start, and keeps three of its five planes: one in general position, one with J negative, and one whose C sits near the +-180 wrap.
• twp-native's timed point-to-point feeds take half a second each instead of one and two.

kins-switch keeps its 10 ms task cycle: at 1 ms its program's G12.1 after a G0 W0 sometimes switches while W is still about 25 um from the end of its move, and W stays there. That looks like a real race in the switch's wait for motion, which I will look into separately.