EKF3 velocity divergence pre-arm

[tridge]

We had a divergence of EKF3 velocity pre-arming on the ground which caused a roll over on takeoff of a quadplane.
GPS velocity looks fine, EKF3 velocity is very bad. Normalised innovations were low.
Replay log here:
http://uav.tridgell.net/tmp/00000082.BIN
replays nicely, no log packet loss
firmware is 4.6.0-beta4, but master replay shows the same issue.

old versions tested:

• 4.5.7 gives same replay result
• 4.4.4 gives the same replay result
• 4.3.8 gives almost the same result (same general VE divergence)
• 4.2.3 reproduces the issue
• the 4.1.x replay doesn't work with the log

you can use replay with EKF2 on this log with this PR:
#29405
and --force-ekf2
Using EKF2 the velocity does not diverge, so this looks to be an EKF3 only issue, and one that has been in EKF3 for a very long time

Apart from the divergence, the other worrying part of this log is that the scaled innovations presented to the EKF failsafe code and to the user over MAVLink do not show any health issues with the EKF. Same goes for the lane switching error scores. So no fail-safes, no GCS warnings, no arming check failures.

Other notes:

• delaying the start of GPS fusion in EKF3 to be after 90s of stability in NavEKF3_core::calcGpsGoodToAlign() instead of 10s "fixes" the issue

[tridge]

V09 is diverging (vertical position variance, note it is using GPS for vertical position)

[tridge]

the V05 (variance of velocity east) is diverging badly:

it is stopping at 1000.0, which comes from NavEKF3_core::ConstrainVariances()

[priseborough]

Replay of the log with EK3_MAG_CAL=4 (always use 3-axis fusion) eliminates the instability so I am now investigating the yaw fusion process.