Turn RunAttributor into a compile-time preference

[vchuravy]

Attributor yet again seems to be the cause of some illegal code transformation,
and I would like my last few days of investigating #3041 (comment) back.

This PR turns running it into an preference flag (so that folks who want it don't need to remember passing an environment flag),
but turns it off by default.

[github-actions]

Benchmark Results

	main	d14a42b...	main / d14a42b...
basics/make_zero/namedtuple	0.0535 ± 0.0018 μs	0.0527 ± 0.002 μs	1.02 ± 0.051
basics/make_zero/struct	0.273 ± 0.0056 μs	0.275 ± 0.0062 μs	0.992 ± 0.03
basics/overhead	5.87 ± 0.01 ns	4.03 ± 0.01 ns	1.46 ± 0.0044
basics/remake_zero!/namedtuple	0.225 ± 0.0075 μs	0.227 ± 0.0087 μs	0.991 ± 0.05
basics/remake_zero!/struct	0.229 ± 0.0094 μs	0.229 ± 0.012 μs	0.997 ± 0.067
fold_broadcast/multidim_sum_bcast/1D	10.2 ± 0.43 μs	10.3 ± 1.8 μs	0.994 ± 0.18
fold_broadcast/multidim_sum_bcast/2D	12.2 ± 0.29 μs	12.2 ± 0.3 μs	0.998 ± 0.034
time_to_load	1.07 ± 0.004 s	1.09 ± 0.028 s	0.982 ± 0.026

Benchmark Plots

A plot of the benchmark results has been uploaded as an artifact at https://github.com/EnzymeAD/Enzyme.jl/actions/runs/25494080677/artifacts/6854668816.

[vchuravy]

Sigh it seems necessary for 1.10 test to pass.

[wsmoses]

is there a reason for doing this since we already have the ref option?

I agree this is nicer, but technically breaking

[vchuravy]

is there a reason for doing this since we already have the ref option?

Attributor is turned off already by default for 1.12 and as #3041 shows causes compilation issues that are really frustrating to debug in 1.11 (I spent ~4days trying to root cause that bug this week).

So yes as CI shows it sometimes does positive things, but it also continuously seems to create deeply frustrating bugs and I for one had enough of it.

[wsmoses]

oh I meant more in reference to it being a preference vs the current ref [not wrt the default or not].

we should look at the remaining failure with 1.11 when disabled. part of why its okay to turn off is ive been slowly making an "attributor lite" pass for our specific needs, though not as aggressive

[vchuravy]

oh I meant more in reference to it being a preference vs the current ref [not wrt the default or not].

I think that's fine. We are promoting it from an internal interface, to something more official and documented.

[wsmoses]

I really worry about making this change as there is a lot of code which modifies the ref atm downstream

[wsmoses]

I also don't like the idea about this being a once-per-session setting

[vchuravy]

I also don't like the idea about this being a once-per-session setting

But it already is, due to caching if you set it inside the session you will get inconclusive results.

With this at least you can setup two projects and change the setting there.