Chore: Benchmaxxing sideffect workflow overhead

[eersnington]

Add workflow overhead benchmarks

Summary

• Add a local benchmark for Sideffect workflow step overhead
• Add coverage for Effect-returning step bodies and rollback handlers
• Cache schema decoder adapters internally without changing the public API

Benchmark Results

Command:

bun run --cwd packages/sideffect bench:workflow-overhead --steps=100000 --samples=3

Results:

case	median ms	delta vs native	us/step
native-fake-baseline	3.696	3.696	0.037
sideffect-direct-struct	60.449	56.753	0.604
sideffect-promise-step	59.496	55.799	0.595
sideffect-effect-succeed-step	73.310	69.613	0.733
sideffect-effect-fn-untraced-step	85.232	81.535	0.852
sideffect-effect-fn-traced-step	219.659	215.962	2.197
sideffect-effect-workflow	61.149	57.453	0.611
sideffect-effect-per-step-wrapper	106.180	102.484	1.062

Notes

• Plain Sideffect struct steps added about 56.8ms over the fake native baseline at 100k local steps.
• Returning Effect.succeed(...) from every step added about 12.9ms over plain Sideffect at 100k steps.
• Named traced Effect.fn("...") in the hot step path was the slowest case at ~220ms.
• These are local fake-step measurements, not Cloudflare production latency numbers.

Verification

• bun run --cwd packages/sideffect bench:workflow-overhead --steps=1000,10000 --samples=3
• bun run --cwd packages/sideffect bench:workflow-overhead --steps=100000 --samples=3
• bun run --cwd packages/sideffect bench:workflow-overhead --steps=1000 --samples=1
• bun run --cwd packages/sideffect test
• bun run --cwd packages/sideffect check
• bun run --cwd packages/sideffect build
• bun run check
• bun run test