Re-calibration on x86 and update cost parameters

[jayz22]

What

Resolves #1246 Also made a minor fix on changing the plain * to saturating_mut for map charges (mentioned in #1141).

This is the first calibration results on x86, previous results were obtained on M1. Overall most of the cost characteristics stay the same, there are some notable changes:

• The VmInstantiation cpu cost has gone down significantly, which explains cpu numbers in most test observations going down (by more than half).
• The arithmetics related cost centers have gone up, this includes {U,I}{256, 128} arithmetics, cryptography functions. This was due to the host observation overhead, it has been fixed.
• The InvokeVmFunction is taking maximum arg size (=32).
• ValSer ValSer has been calibrated using closer-to-worst examples

The mem-related cost parameters are using analytical inputs (not calibrated), same as before.

raw output

Why

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

Known limitations

[jayz22]

Thanks @dmkozh and @graydon for the review. I've refreshed the PR again after your comments. Now it should be ready to go again.

One thing that's been updated since your last review was the calibration of xdr conversions (ValSer, ValDeser), previously we were was treating it like a mem op (using the same analytical numbers from memcpy). But that's because previously it was calibrated to a not-too-worst case input, which was just (de)serializing a large bytes array. I've since updated the samples to deeply nested structure trying to resemble the worst case, and have updated the parameters accordingly. The cpu cost parameters are directly taking from calibration, the memory cost parameters are still analytical but based on calibration result (~a factor of 3 bytes per input byte). I think this is the best we can do but curious what you guys think.

[graydon]

Great work, thanks!