motion: fix the split read check on the shared status structs

[grandixximo]

Task reads emcmotStatus, emcmotConfig and emcmotInternal by copying the struct and keeping the copy when the copied head equals the copied tail. head is copied first and tail last, so a copy that straddles the start or end of a motion write is kept torn. Only copies wholly inside a write were rejected.

I measured it with markers written at both ends of every status write: about 0.2% of the copies task kept were torn, on x86 locally and on the GitHub x86 and ARM (Neoverse-N2) runners.

Fix:

• New rtapi_seqlock.h: read tail before the copy and head after it, with release/acquire ordering. On ARM the right pairing alone still kept torn copies; the fences are needed.
• All writers (controller, command handler, emcmot_config_change(), motion-logger) and the three readers in usrmotintf.cc use it.
• head/tail widen from unsigned char to unsigned int, so a reader stalled mid-copy cannot see the counter wrap back to the same value (256 writes, about 256 ms at a 1 kHz servo thread). They stay plain fields so the structs stay trivially copyable in C++.
• Writes that bypassed the counters: SET_NUM_SPINDLES, SET_MAX_FEED_OVERRIDE, SETUP_ARC_BLENDS, SET_PROBE_ERR_INHIBIT and SELECT_KINS_TYPE changed emcmotConfig without emcmot_config_change(), so task also never re-read them; the controller wrote emcmotInternal (the planner included) without a window; emcmotSetCycleTime() ran before the controller's window opened.

A read now needs a moment with no write under way, long enough to copy the struct (about 6 KB). If the servo thread overruns continuously and nothing runs between motion-controller and the next motion-command-handler, task gets no status at all, where the old check got torn copies. Other HAL functions after motion-controller, as in any config with hardware, leave that gap.

Test: tests/rtapi-seqlock runs a writer thread against rtapi_seqlock_read() for 3 s and fails on any torn copy kept.

	x86	ARM
old protocol, 3 s	18k-23k torn kept	75k-190k torn kept
right pairing, no fences, 30 s	0	8.7k-2.3M torn kept
this PR, 30 s	0 of ~32M	0 of ~37M
motion sim with markers, this PR	0 torn of 6.1M	0 torn of 6.2M

runtests -u passes locally.