RSDK-12719: Lower the number of arm positions we cache.

[dgottlieb]

Starting a conversation via a PR. Apologies if this isn't the best place.

We got a report of crashing once upgrading v101/v102. This was a pretty clear OOM on the first access of motion planning on a lite6 arm (read: building the cache). This was on one of the nanos I think? With 8GB of RAM that's shared with the GPU. I believe a CV model was also loaded.

The reason for this increased consumption was the finer grained jog creating more cache entries. Here are some numbers for memory used by the cache (after forcing a GC) across different versions (Alloc being actively used memory):

v0.102.0
Alloc = 2861 MiB	TotalAlloc = 10051 MiB	Sys = 5543 MiB	NumGC = 23

v0.101.0
Alloc = 2379 MiB	TotalAlloc = 7350 MiB	Sys = 4714 MiB	NumGC = 20

v0.100.0
Alloc = 253 MiB	TotalAlloc = 1756 MiB	Sys = 619 MiB	NumGC = 19

On main which removed the unused pose memory was cut by ~35% from v101/v102. But still greatly increased from v100:

Main:
Alloc = 1648 MiB	TotalAlloc = 5170 MiB	Sys = 3785 MiB	NumGC = 13

This patch brings memory back in line with the v100:

arm6JogRatios   = []float64{90, 32, 8, 8, 4, 2}
Alloc = 227 MiB	TotalAlloc = 1376 MiB	Sys = 562 MiB	NumGC = 15

[dgottlieb]

An "enhanced" version of this patch is to query how much memory the machine has. Such that we're more conservative with low-memory rpis while getting better precision with bigger boxes.

Happy move this PR in that direction. Or any other direction. Just wanted to get product feedback/a path forward for users unable to upgrade.

[erh]

Who can't upgrade?
This will make a lot of things worse and isn't the direction I think we should go.
We can just not cache at all if the system has less than x amount of ram perhaps

[dgottlieb]

Nick H here: https://viaminc.slack.com/archives/C091QGG4MQ8/p1763500966092209

[erh]

lame - is it easy enough to see available memory?

[github-actions]

Availability

Scene #	viamrobotics:main	dgottlieb:RSDK-12719	Percent Improvement	Health
1	100%	100%	0%	➖
2	100%	100%	0%	➖
3	100%	100%	0%	➖
4	100%	100%	0%	➖
5	100%	100%	0%	➖
6	100%	100%	0%	➖
7	100%	100%	0%	➖
8	100%	100%	0%	➖
9	100%	100%	0%	➖
10	100%	100%	0%	➖

Quality

Scene #	viamrobotics:main	dgottlieb:RSDK-12719	Percent Improvement	Probability of Improvement	Health
1	1.31±0.00	1.31±0.00	0%	63%	➖
2	0.90±0.00	0.90±0.00	-0%	50%	➖
3	5.86±1.18	5.14±0.00	12%	73%	➖
4	3.13±0.40	3.23±0.42	-3%	43%	➖
5	10.26±3.30	9.47±3.34	8%	57%	➖
6	9.67±3.46	10.51±4.32	-9%	44%	➖
7	6.71±2.36	4.89±2.29	27%	71%	➖
8	0.90±0.00	0.90±0.00	-0%	50%	➖
9	4.15±0.00	4.20±0.14	-1%	37%	➖
10	12.84±0.41	12.84±0.41	-0%	50%	➖

Performance

Scene #	viamrobotics:main	dgottlieb:RSDK-12719	Percent Improvement	Probability of Improvement	Health
1	0.03±0.01	0.03±0.01	3%	54%	➖
2	0.04±0.01	0.03±0.01	20%	82%	➖
3	0.09±0.03	0.03±0.00	67%	97%	✅
4	1.19±0.04	1.17±0.05	2%	62%	➖
5	1.70±0.38	1.62±0.35	5%	56%	➖
6	1.81±0.57	1.93±0.61	-7%	44%	➖
7	2.18±0.75	2.07±0.75	5%	54%	➖
8	0.03±0.00	0.03±0.00	11%	78%	➖
9	2.04±0.14	1.94±0.07	5%	73%	➖
10	3.36±0.54	3.34±0.56	0%	51%	➖

The above data was generated by running scenes defined in the motion-testing repository
The SHA1 for viamrobotics:main is: b39abd91b2d7c568703e39f08cf352167db9a609
The SHA1 for dgottlieb:RSDK-12719 is: b39abd91b2d7c568703e39f08cf352167db9a609

• 10 samples were taken for each scene