Benchmark resource cost per entrypoint and fail CI on regression

[joelpeace48-cell]

Problem

Soroban meters CPU instructions and memory against a per-transaction budget.
Exceeding it makes an entrypoint permanently uncallable — for a deployed
contract, that is a bricked function with no fix short of an upgrade. Nothing
currently measures how close any entrypoint is to that ceiling.

What to do

• Use env.budget() in tests to capture CPU and memory per entrypoint.
• Commit a baseline and fail CI when a change exceeds it by more than a small
  threshold.
• Measure at realistic input sizes, especially for batch operations — cost is
  usually linear in a way that only shows up at scale.
• Publish the numbers in the README, alongside the WASM artifact sizes.
• Track headroom explicitly: percentage of budget consumed, not just absolute
  numbers.

Acceptance criteria

• Budget measured per entrypoint at realistic input sizes
• Baseline committed; CI fails on regression beyond threshold
• Headroom reported as a percentage of the limit
• Numbers published

Notes

Do this before the contracts grow. Discovering an entrypoint is at 95% of
budget after deployment leaves no room to fix a bug in it.