Precompile AIR count reduction

The precompile proof currently has [ten AIR instances](https://github.com/0xMiden/miden-vm/blob/18d9f6361acbf2f8fb22e4780b312e4b4f96cbce/crates/precompiles-air/src/air.rs#L27-L45). Combine compatible AIRs so the proof has about five instances. A likely layout keeps Keccak and the byte-pair lookup separate. It stacks the remaining tables by hash or transcript work, uint work, and curve work.

The issue is complete when:

- the design names each combined AIR and its row selector.
- the prover and verifier use the new canonical instance order.
- bad-trace tests cover every selector and cross-table lookup.
- benchmarks report proof size and peak memory. They also report proving time.
- the final design reaches about five AIRs or records why another count performs better.

[#3417](https://github.com/0xMiden/miden-vm/issues/3417) only repacks lookup columns and remains separate.