LTO: Use -flto and -flto-partition only as needed

[dhalbert]

-flto slows down a build by a factor of two or so. We were also always using -flto-partition=none, which saves more code space than the default -flto-partition=balanced. However, one some builds that need -flto, they don't really need the extra savings of -flto-partition=none. Using the default -flto-partition reduces build times to 2/3-3/4 of the original.

I did a bunch of experiments to see what space savings were lost, and also tried the other -flto-partition= options.

Also I tried -flto=auto, which does parallel make, but that seems to be the default. Leaving it off did not make any difference.

Here are some Metro M4 builds, done with -j12 on my Intel i7-8700T Linux dev system. The M4 builds are generally -O2, but need LTO for that to fit. Times rounded to half seconds, because the variance between runs can be about that. The first line in each is a clean en_US build; the second is a pt_BR build in the same build dir as en_US, to avoid unnecessary recompiles, as is done in the GitHub Actions.

-flto-partition=	firmware bytes free	build time (secs)	translation
1to1	32316	29	en_US (clean)
1to1	29568	16.5	pt_BR (using en_US build)
balanced	32296	28	en_US (clean)
balanced	29556	14.5	pt_BR (using en_US build)
none	33484	40	en_US (clean)
none	30728	27	pt_BR (using en_US build)
one	33684	40	en_US (clean)
one	30928	27	pt_BR (using en_US build)

So flto-partition=one is slightly better than -flto-partition=none, and the time is the same. balanced is much faster than than none or one, as you can see. 1to1 is worse than balanced.

I also did some trials on Trinket M0 and CPX. one is very slightly better than none (8-18 bytes). one vs balanced is 92 bytes better for Trinket M0 and 120 bytes better for CPX. So I left -flto-partition-=one on for all SAMD21 builds, and for CIRCUITPY_FULL_BUILD=0 SAMx5x builds.

We only use LTO now for atmel-samd and nrf builds. In this PR, I turned off LTO on roomy nrf builds, which saves a lot of time. Roomy is very roomy - there are still 100k's of unused space.

I'll look at the total time savings after the CI runs.

[dhalbert]

This is ready for review. The single build failure appears to be a CI problem.

[tannewt]

Thank you!

[dhalbert]

It's a little hard to tell, because the build duration times include waiting time, but I think this is saving about 15 minutes, reducing a full build from about 80 minutes to 65 minutes.