feat(avm): set avm circuit subgroup size

[fcarreiro]

I had to make a few changes in this PR, so bear with me.

Inside you there are two wolves:

• The size of the generated trace (which I'll call the "trace"): e.g., you run a loop from 1 to 1000 and get a trace with size 1000 (or, if you will, ~2^18 if you add the precomputed columns)
• The size of the polynomials used for proving: aka circuit subgroup size. We need to set this to some number, since the VK and other things depend on it.

In this PR I'm setting the latter to 2^21. Is that all? No, because this still needs to work with traces < 2^21. Can't you just resize the trace to 2^21 and call it a day? You can, but then your memory and time will suck.

This PR therefore does the following: Suppose your trace has size 1000 and you know already our subgroup size is set to 2^21. The polynomials will be initialized with a real size of 1000 rows, and a virtual size of 2^21. Then the values will be set from our generated trace.

• This is far better in terms of memory because you only "pay for what you use"*.
• This is also far better in terms of time, because resizing the trace to 2^21 takes forever (like 20s+). This is because we currently use std::vector which forces the initialization of every field, even if you previously reserved memory (which btw is fast).

Extra: I also did some cleanups, in particular I try to rely less on environment variables and have a clear flow separation between "prod" and tests.

• bb avm_prove only runs check circuit if you really ask for it.
• bb avm_prove uses full proving by default (all range checks and precomputed tables). This will in particular help with a more realistic devnet/testnet.
• tests manually set the above options
• check-circuit only checks rows up to the "trace" size; this should make it faster and still sound.
• the 2^21 size does not (effectively) affect check-circuit

Results: I'm running this program, which at 2^22 rows took 6 minutes and 280GB ram. Let's then suppose that for 2^21 it would've taken 3 minutes and 140GB ram.

• After this PR, proving takes 49 seconds, and 31GB ram. (note that the time gains include as well the last few PRs)

*that is, if you use 1000 rows, you allocate 1000 rows. Sparcity is not yet taken into account. We need some more changes for that.