Host budget metering, cost model, calibration changes

[jayz22]

What

This is a set of sweeping changes/fixes/cleanups to host calibration and metering:

1. Clean up of cost types:

• Add HostMemAlloc/Cpy/Cmp for host memory allocation, copy, and comparison
• Remove redundant cost types or consolidate them with more general types such as HostMem types, e.g. ScVecToHostVec, CloneEvents, HostObjAllocSlot, and the single byte cost types i.e. BytesPush/Pop/Insert/Del.

2. Clean up unused model terms: log, quadratic

3. Add a "baseline" calibration and subtract it from the actual measurements for more accurate results

4. Remove sample deallocation (using std::mem::forget) from the measurement results for better accuracy.

5. Various fixes:

• ChargeBudget was previously not increasing the input
• MeteredVector, MeteredMap when to charge_new.
• metered_scan_slice_of_slices
• test_declared_size was being too strict (using assert_eq instead of le), resulting in failure in CI but passes locally.

6. Changes for improving debug experience

• Add per cost per CostType to the Budget, add Debug impl for it.

7. Update the cpu and mem cost parameters based on the newly calibrated results.

Below are the calibration output.
output6.txt

Why

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

Known limitations

[TODO or N/A]

[dmkozh]

I haven't fully reviewed this yet, 11 files are still remaining.

[graydon]

Generally good! Very glad to see:

• Number of cost types shrinking, eliminating duplication and ambiguity
• Smaller, simpler and more precise calibration functions
• Simpler models
• Actual calibration results that are closer to reality

Handful of nits noted in review, and some questions and concerns, broadly speaking:

• Might want to avoid the forget calls in favour of a pass/return protocol
• Might want to black_box some stuff
• Some of the byte-growth metering seems like it's overcharging
• Calibration numbers seem like a few have noise in them or don't make sense
• A few calibration functions don't seem like they're counting exactly what they should

[jayz22]

@graydon I have just pushed three commits that addresses your major concerns:

• Use recycling patterns instead of std::mem::forget
• Putting runner inside std::hint::black_box to prevent unexpected optimization.
• Resolve bytes over-charging.
  I think this should solve all of the structural issues. I will go on to address the remaining "nit"/one-off comments.

[graydon]

@jayz22 ok I've marked most of the issues I had as resolved, and added a few very minor nits around saturating adds. I think mostly I'm good with this now, though there's still a bunch of commentary from @dmkozh to resolve as well. I'm happy to merge it whenever ready (and I concur it'd be nice to get into this release, which is starting tomorrow, so if you and @dmkozh could give it some focused time this afternoon maybe we can make it?)