Add synthetic-transaction regression bench

[Al-Kindi-0]

Purpose:

Adds a new benchmark crate under benches/synthetic-tx-kernel/ that reproduces the proving-cost brackets of a real Miden transaction from a small JSON snapshot, without pulling in any protocol-level code. The result is a VM-level regression detector that isolates prover changes from workload changes: when you change the VM, the bench moves; when protocol changes its transaction kernel, miden-vm's bench stays put until a new snapshot is explicitly refreshed.

This is a simplified variant of the approach originally prototyped in 0xMiden/protocol#2391. The key simplification: instead of maintaining a fused cross-repo benchmark that depends on protocol's transaction machinery at runtime, we row counts and emit an equivalent synthetic MASM program on the miden-vm side. The two repos stay loosely coupled via a hand-maintained schema, and miden-vm's bench has no runtime dependency on protocol.
The snapshot producer lives in a companion protocol PR (bin/bench-transaction/src/bin/tx-trace-snapshot.rs).

How it works

Per bench run:

1. Calibrate each MASM snippet's per-iteration cost against the current VM (once).
2. For each snapshot JSON: solve for per-snippet iteration counts via fixed-point refinement, emit a begin ... end MASM program, verify it lands in the target padded-trace brackets (padded_core_side and padded_chiplets), then run execute + execute_and_prove Criterion groups.

A bracket miss fails the bench; sub-bracket drift is reported but tolerated because proving cost is driven by padded trace length, not raw row count.

Validations

Wall-clock prove times against protocol's real bench-transaction prove benches:

scenario	protocol (real)	synthetic (miden-vm)	delta	% diff
consume-single-p2id	970 ms	998 ms	+28 ms	+2.9%
consume-two-p2id	1886 ms	1888 ms	+2 ms	+0.1%

[huitseeker]

Thanks a lot, this looks great! I added a commit that improves how the snapshot test catches wrong, missing, or extra committed JSON files.