Setup bench framework to run experiments

[jayz22]

What

Setup bench framework to run experiments

• Experimental bench code is located in experimental (benches/common and src/cost_runner dirs)
• Aggregate ContractCostType, WasmInsnType and a new ExperimentalCostType into top level CostType

Resolves #1030

• Bring back EdwardsPointCurve25519ScalarMul as a experimental cost type and run variation histogram on it.

Why

[TODO: Why this change is being made. Include any context required to understand the why.]

Known limitations

[TODO or N/A]

[jayz22]

Calibration result shows both EdwardsPointCurve25519ScalarMul and VerifyEd25519Sig have a large variation (max/min = 14 and 6 respectively, see screenshot), which confirms @graydon's concern in #1030

Will next figure out how to address it.

[jayz22]

Upon closer examination, I change my conclusion above. The worst case vs the average case is actually pretty close.

I realize the above experiment isn't really testing what we are looking for. We are trying to answer the question "is our calibration, which is based on the average case, underestimating the worst case".
Here are some tweaks to the setup in order to better answer that question:

1. Throw away the "best" case. For EdwardsPointCurve25519ScalarMul (doesn't matter too much for VerifyEd25519Sig ), the best case is significantly faster than the average, but it's rare and doesn't contribute to the problem statement above.
2. Fix the input size. For VerifyEd25519Sig that is the message size. The variation analysis should fix the input size and examine the effect of random bytes in the input. (Longer message will cost more to verify sig but that's not what we are interested in discovering)
3. Increase the iteration count from 100->1000

With the above tweaks, there are the results:

The ration between worst/best case for EdwardsPointCurve25519ScalarMul is 1.34.
For VerifyEd25519Sig there is no trivial way to set the worst case, so everything is random. With 1000 sample points, the ratio between max/min is just 1.008.

So I conclude the worst case of ed25519 is not significantly worst than the average, thus no need to tweak our calibration setup or numbers.

[graydon]

Looks good! Thanks for investigating