perf(target): enable unaligned-scalar-mem for riscv64im guests

[diegokingston]

Summary

Adds +unaligned-scalar-mem to the riscv64im-lambda-vm-elf target spec. This lets LLVM lower potentially-unaligned u64 accesses (align-1 loads/stores, e.g. u64::from_le_bytes on byte buffers in the keccak absorb path, ptr::read_unaligned) to single ld/sd instructions instead of byte-assembled lbu/slli/or and sb/srli chains.

Sound on this target: the VM tolerates unaligned doubleword accesses — executor/src/vm/memory.rs load_doubleword/store_doubleword byte-assemble them host-side at the same 1-instruction cost as aligned ones.

Measured guest cycles (cli execute --cycles, deterministic)

Fixture	Baseline	Patched	Delta
ethrex_empty_block	994,639	763,430	−23.3%
ethrex_simple_tx	1,791,453	1,445,399	−19.3%
ethrex_10_transfers	6,810,436	6,200,610	−9.0%

The % shrinks with tx count as keccak/hashing becomes a smaller share of total cycles.

Verification

• Codegen: keccak provider absorb loop goes from 8×lbu+slli/or per lane to one ld per lane (disassembly of the built ethrex guest, before/after in commit body).
• Minimal no_std PoC (*lane ^= u64::from_le_bytes(buf), read_unaligned, write_unaligned): baseline emits byte chains, patched emits single ld/sd.
• cargo test -p executor --test rust test_keccak passes with a patched guest ELF; keccak guest cycle count unchanged (aligned workload).
• CI note: ELF artifact cache keys in .github/workflows/pr_main.yaml hash the target JSON, so cached guest ELFs invalidate naturally.

[diegokingston]

/bench

[diegokingston]

/bench

[github-actions]

Benchmark — ethrex 20 transfers (median of 3)

_{Table parallelism: auto (cores / 3)}

Metric	main	PR	Δ
Peak heap	57086 MB	51336 MB	-5750 MB (-10.1%) 🟢
Prove time	30.840s	27.184s	-3.656s (-11.9%) 🟢

🎉 Improvement detected — heap or time decreased by more than 5%.

✅ Low variance (time: 1.6%, heap: 3.9%)

_{Commit: c4e42b2 · Baseline: cached · Runner: self-hosted bench}