ci: Add comparative GPU benchmarks on merge_group

[samuelburnham]

GPU benchmarks for PR regression testing

• Adds a benchmark regression test when merging a PR, based on https://github.com/lurk-lab/lurk-rs/blob/main/.github/workflows/merge-tests.yml
• Refactors the recursive-snark benchmark group and ID to enable formatting with criterion-table. This will be extended to compressed-snark and their supernova counterparts when ready.
• I attempted to test the criterion-table output in a621594#commitcomment-136557829, and the basic logic works. Unfortunately, we're limited by the fact that the base benchmarks on dev don't have the same formatting or parameters as this PR, so it's impossible to demo exactly what it will look like beyond the proposal below.
  • It does correctly pick up the artificial regression I introduced, which opened 🚨 Performance regression in #244 #245 (partially thanks to the Dockerfile fix in Install bc on GPU runners argumentcomputer/gh-actions-runner#25)
• I also can't run these actions in a fork, so rather than a suboptimal reconstruction in https://github.com/lurk-lab/ci-lab I'd like to merge this PR and then iron out any last issues in the interests of speed.

Proposed criterion-table output:

Benchmarks

Table of Contents

• Benchmark Results
  • RecursiveSNARK

Benchmark Results

RecursiveSNARK

	ref=ae521be709871f94a622c85b0e5d6e8f776be144	ref=ace6724dc77667506dbcd1122b9851bd6edb25d2
Prove-StepCircuitSize-0	29.28 ms (✅ 1.00x)	28.71 ms (✅ 1.02x faster)
Verify-StepCircuitSize-0	19.48 ms (✅ 1.00x)	19.10 ms (✅ 1.02x faster)
Prove-StepCircuitSize-6559	35.57 ms (✅ 1.00x)	35.27 ms (✅ 1.01x faster)
Verify-StepCircuitSize-6559	23.84 ms (✅ 1.00x)	23.89 ms (✅ 1.00x slower)

---

Made with criterion-table

Design notes

For the other benchmarks which have more than one group, they would each appear in their own table (e.g. CompressedSNARK and CompressedSNARK-Commitments). This layout will be shown as well when implemented.

Some other display options:

• Split each StepCircuitSize into its own group, which is the existing behavior. This would proliferate the number of benchmark tables to a large degree if we are displaying four different sets of benchmarks in one .md file
• Split the Prove and Verify benchmarks into their own table or group them together. This would require some finagling since the benchmarks are currently ordered sequentially in the order they are run.

Future work

• Refactor into a reusable workflow (Add a working comparative GPU benchmark workflow/action argumentcomputer/ci-workflows#15)
• Cache the PR benchmark for reuse (CI: Cache comparative GPU benchmarks #246)

Helps #243

[huitseeker]

Agreed, this looks like a lot of progress, let's merge and iron out the kinks.