[release-0.5] Add patch to set -O2 for suitesparse build

[staticfloat]

This fixes the compiler regression noticed in GCC 6.3.0 in #20123 by disabling -ftree-slp-vectorize as a side-effect of moving from -O3 to -O2. We prefer to compile with -O2 by default anyway, so let's do so.

[tkelman]

as discussed in #20165, more of our makefiles default to -O3 than -O2, so would rather this be targeted to the architectures where it causes a problem, or be demonstrated that it actually doesn't make a difference in a test that thoroughly exercises suitesparse

[ViralBShah]

It wasn't right for our Makefiles to use -O3 in the first place. The policy should be to use -O2 and -O3 in a targeted way, which is a well established practice. I am ok with merging this right away.

[tkelman]

Not without a targeted test of the performance impact. It's not acceptable to slow down sparse linear algebra for everyone to avoid a compiler bug on a rarely used architecture.

[staticfloat]

The distro packaging wisdom of Debian, Ubuntu, and Arch all agree that O2 should be the default. However, we are much more performance-oriented than these packages. So I think we should move over to O2 being the default unless there is a significant performance impact. I will measure the performance impact across a few different platforms, if you can come up with a good benchmark, Viral. I'm not sure we have a good suitesparse benchmark that we can use to measure speed, or if we do which parts are the most important.

[tkelman]

I'm going to close this so we can discuss in a single place, the master PR #20165, until we've shown one way or another whether -O3 is making a difference.