refactor(prover): centralize u64 limb decomposition into a limbs helper

[diegokingston]

What

Centralizes the repeated u64 → 16/32-bit little-endian limb decomposition that was open-coded across the VM trace tables. Adds a small prover/src/tables/limbs.rs module and converts ~20 tables to use it.

The helper (prover/src/tables/limbs.rs)

pub const fn limbs_16(x: u64) -> [u64; 4]   // [x[0..16], x[16..32], x[32..48], x[48..64]]
pub const fn limbs_32(x: u64) -> [u64; 2]   // [lo, hi]
pub const fn limb_16(x: u64, i: u32) -> u64 // (x >> 16*i) & 0xFFFF
pub fn set_limbs_16(data: &mut [FE], col: usize, x: u64) // FE limbs into data[col..col+4]
pub fn set_limbs_32(data: &mut [FE], col: usize, x: u64) // FE limbs into data[col..col+2]

Tests live in the dedicated prover/src/tests/limbs_tests.rs (no inline tests in the production module).

Conversions

set_limbs_16 / set_limbs_32 replace the data[base + cols::X_0] = FE::from(v & 0xFFFF); … blocks; limbs_16 / limbs_32 replace the integer [u32;4]/[u64;…] decomposition arrays. Tables touched: mul, dvrm, decode, cpu, cpu32, commit, load, page, register, memw, memw_aligned, memw_register, lt, branch, eq, store, ec_scalar, ecdas, keccak, keccak_rnd, shift, plus the trace_builder bus-value arrays.

Every converted column group was verified consecutive in its cols module before conversion.

Deliberately left as-is

Parametric column indices (cols_i[..], state_ptr(lane, k), lo_col/hi_col vars), byte (& 0xFF) / nibble (& 0xF) splits, non-standard groupings (e.g. DWordWHH, limbs starting at index 1), sign-fill constants, and the local half closure in trace_builder (already a local dedup).

Correctness

Pure refactor — byte-identical trace output. cargo test -p lambda-vm-prover --lib is unchanged at 322 passed / 104 failed; the 104 failures are pre-existing and environmental (tests that read executor/program_artifacts/*.elf, absent in the sandbox — they fail identically on main). cargo clippy -p lambda-vm-prover is clean. The 5 new limbs_tests cover the helper (LE order, reconstruction, a drift-guard against the open-coded form, and consecutive-column writes).