Add fabs support to Taylor series expander

[pavpanchekha]

This PR adds support for fabs in the Taylor expander. It thus fixes #1271. Specifically, (fabs (+ x 1)) expands to (+ 1 x); (fabs (- x 1)) expands to (+ 1 (neg x)) (it gets negated because the constant term is positive); and finally, terms like (fabs (+ x (* x x)) expands to (+ (fabs x) (* (fabs x) x)) which is confusing but accurate in a neighborhood near 0.

Initially I had copilot write it and I was impressed that it came up with something plausible, but actually its code didn't work at all, so I had to rewrite from scratch basically. https://chatgpt.com/codex/tasks/task_e_684b1775fec4833193db8dcc2a906e7a

This PR also includes a tiny amount of cleanup for trigonometric stuff, leveraging new helper function I added for fabs.

[pavpanchekha]

I can confirm that this fixes #1271, at least on my machine.

[pavpanchekha]

Unfortunately this seems to hurt accuracy a bit, branch vs main. I'll try to look into it. The effects aren't huge but not worth it just for less weird output.

[AYadrov]

I slightly misunderstand how the code works, but it looks fine to me

[pavpanchekha]

The biggest drop is "a from scale-rotated-ellipse" but it's kind of impossible to tell what's going on there.

[pavpanchekha]

Looking a bit more I realize that this branch wasn't up to date with main, so the comparison might not be apples-to-apples, will look again tomorrow morning.

[pavpanchekha]

Ok, I looked at a couple of cases and they are all cases of over-greediness. Basically, the ability to Taylor-expand more stuff means that Herbie achieves better results after iter 1, but that only prevents it from searching harder for a rewrite-based solution.