Actually use precomputed witness

[arthurpaulino]

This PR is a second attempt after the failure of #928 (reverted in #931).

#928 was returning the wrong output because LEM doesn't allocate the output at the end.

So this PR allows MultiFrameTrait implementers to cache their witnesses internally, using whichever data format they want. LEM's MultiFrame does so by caching the WitnessCS and the output (the vector of allocated numbers).

With that available, LEM can now make proper use of cached witnesses.

[arthurpaulino]

!gpu-benchmark

[github-actions]

Benchmark for 08da70d

Click to view benchmark

Test	Base	PR	%
LEM Fibonacci Prove - rc = 100/fib/num-100	4.0±0.01s	3.9±0.02s	-2.50%
LEM Fibonacci Prove - rc = 100/fib/num-200	8.1±0.02s	7.8±0.02s	-3.70%
LEM Fibonacci Prove - rc = 600/fib/num-100	3.4±0.01s	3.3±0.01s	-2.94%
LEM Fibonacci Prove - rc = 600/fib/num-200	7.6±0.04s	7.4±0.11s	-2.63%

[arthurpaulino]

!gpu-benchmark

[github-actions]

Benchmark for 08da70d

Click to view benchmark

Test	Base	PR	%
LEM Fibonacci Prove - rc = 100/fib/num-100	5.0±0.02s	4.8±0.04s	-4.00%
LEM Fibonacci Prove - rc = 100/fib/num-200	10.2±0.04s	9.9±0.03s	+-2.94%
LEM Fibonacci Prove - rc = 600/fib/num-100	4.9±0.02s	4.8±0.03s	-2.04%
LEM Fibonacci Prove - rc = 600/fib/num-200	10.7±0.03s	10.7±0.23s	0.00%

[arthurpaulino]

!gpu-benchmark

[github-actions]

Benchmark for 86b5844

Click to view benchmark

Test	Base	PR	%
LEM Fibonacci Prove - rc = 100/fib/num-100	5.0±0.03s	4.7±0.03s	-6.00%
LEM Fibonacci Prove - rc = 100/fib/num-200	10.1±0.04s	9.7±0.03s	+-3.96%
LEM Fibonacci Prove - rc = 600/fib/num-100	4.8±0.02s	4.7±0.03s	-2.08%
LEM Fibonacci Prove - rc = 600/fib/num-200	10.5±0.03s	10.2±0.05s	-2.86%