Store top of the stack in memory channel 0

[hratoanina]

Updated version of #1196. This one only introduces one extra column (by reintroducing the get_context and set_context flag).

This PR stores the current top of the stack in memory channel 0, saving memory operations in a lot of cases.
Reading and writing the new top of the stack after an operation is straightforward, but the most annoying thing was a disjunction depending on whether the current stack was empty or not for pushing instructions (resp. the next stack was empty or not for popping instructions), and whether or not to disable a channel in that case. This was done with prover-provided inverses which could be stored in general CPU columns, thus not adding any new columns.

Further work:
I couldn't get to mix this PR with the column merging while keeping stack constraint degrees below 3. I have an inkling this should be possible with more general columns though.

closes #1149

[hratoanina]

Some figures about the memory trace reduction:

add11
517719 -> 440537 --> 14.9% reduction

(on a slightly outdated branch but still relevant)
pairingTest_0
6916972 -> 5498226 --> 20.5% reduction
ecmul_0-0_0_21000_0
1654468 -> 1443551 --> 12.7% reduction
ecadd_0-0_0-0_21000_0_0
1654939 -> 1443587 --> 12.8% reduction
CallSha256_0
883714 -> 769612 --> 12.9% reduction
CALLCODEBlake2f_0
509411 -> 439444 --> 13.7% reduction

[hratoanina]

Failures are from interpreter tests, since I forgot to update the top_stack logic there, but other integration tests pass. I'm working on fixing this.

[hratoanina]

@wborgeaud
It's a big one, sorry about that! Don't hesitate to mention anything about the design choices, there were a lot of arbitrary ones.

[Nashtare]

Not an actual review, just quickly skimmed through. Only the comment on decode.rs matters though, the rest is nitpicking cleanup.

[nbgl]

This is excellent work!! Thank you.

I'm sorry it took me so long to review this.

[hratoanina]

Thanks for the review! We're not merging right now since we found some tests from the EVM test suite which aren't passing.