perf(syscalls): zero-copy private input API and migrate guests

[diegokingston]

Summary

get_private_input() used to end in slice.to_vec(), bulk-copying the whole private input every run even though the region is a fixed static mapping (0xFF000000) and a zero-copy path already existed (ef_io::read_input). This changes the API to return &'static [u8] and migrates the in-repo callers:

• syscalls/src/syscalls.rs: return the from_raw_parts slice directly (host writes the mapping before execution and never mutates it, so the borrow is valid for the whole run); doc + SAFETY notes updated.
• Guests migrated: ethrex (rkyv::from_bytes(input)), commit, commit_sum. memory needed no change (its try_into() resolves via TryFrom<&[u8]>).
• syscalls/README.md API table updated.

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

Fixture	Baseline	Patched	Delta
ethrex_empty_block	994,639	980,441	−1.4%
ethrex_simple_tx	1,791,453	1,723,120	−3.8%
ethrex_10_transfers	6,810,436	6,743,623	−1.0%

Savings exceed the single-syscall-copy model (~3 instrs per 8 input bytes): the owned Vec also forced downstream clones in the guest's rkyv deserialization layers, which are gone too.

Verification

• riscv64 guest-target cross-checks: lambda-vm-syscalls, commit, commit_sum, memory, and the full ethrex guest — all clean.
• No host-workspace crate uses this API (crypto/ethrex-crypto uses only ecsm_mul/keccak_permute).

Follow-up (not in this PR)

• bench_vs/lambda/recursion (untracked WIP guest) needs the same call-site tweak (postcard::from_bytes(&blob) → (blob)) once this lands.
• Pre-existing repo quirk surfaced while verifying: syscalls/ is not in the root workspace's exclude, so cargo check from its own dir errors ("believes it's in a workspace when it's not"); an empty [workspace] table in its Cargo.toml would fix it.

[diegokingston]

Closing in favor of the zero-copy private-input implementation already on feat/hint-table (ead3da8a — additive get_private_input_slice(), no API break). Measured numbers in this PR's body still hold for that approach (−3.8% cycles on ethrex 1-tx, ~66k/run).