Trace-gen cleanup: single MU-column source, simpler push_reg_access, tighter visibility, stale docs

[MauroToscano]

Cleanup bundle from #786's final review (findings #1/#3/#4/#7/#8). No behavior change; prove+verify pass byte-identical.

#1 — one source of truth for the MU-column map (bitwise.rs)

The compile-time guard proved ALL ↔ lookup_type_index is a bijection but did not cover mu_column itself, so a future typo mapping two lookup types to the same MU column would be silently lossy under fill_multiplicities' overwrite semantics (and diverge from update_multiplicities, which increments). Now a single const MU_COLUMNS: [usize; NUM_LOOKUP_TYPES] in lookup_type_index order is the only type→column source; mu_column/type_mu_column both read from it, and a const _ assert checks the entries are pairwise distinct — a duplicate column is now a build error. Also deletes the second 10-arm match.

#7 — simpler push_reg_access (trace_builder.rs)

Was 8 loose scalars (val0/val1/old0/old1 among them) plus a build_fallback closure repeated at 3 call sites, needing #[allow(clippy::too_many_arguments)]. Now takes val: [u32;2], old: [u32;2] (kills the transposition footgun — a swapped arg would silently write a wrong trace cell) and a shared build_reg_fallback builds the op internally. Verified byte-for-byte against the old closures: reads pass val==old==reg_value, the write passes val=reg_value, old=old_value; the fallback packs [v0,v1,0,…] which is identical since pack_register_value already zeroes indices 2–7.

#8 — tighter visibility (bitwise.rs, memw_register.rs)

BitwiseHistogram + its methods and the MU-column/lookup-type helpers (mu_column, lookup_type_index, NUM_LOOKUP_TYPES, BitwiseOperationType::ALL) are internal to the prover crate → pub(crate) (grep-verified no cross-crate use). The MemwOperation-based generate_memw_register_trace had only test callers → #[cfg(test)] (production uses generate_memw_register_trace_from_rows).

#3, #4 — stale docs

MemwSink's Vec impl is used by the count_table_lengths sizing pass, not "tests/scratch"; generate_bitwise_trace's MU columns are filled by fill_multiplicities, not update_multiplicities (which now only tops up the continuation L2G path).

Testing

cargo test -p lambda-vm-prover --lib: 502 passed (only the 5 known missing rust-guest ELF fixtures fail, at file-read). make lint, clippy --all-targets, fmt --check clean.

Independent of #794 (the parallelism ABBA PR); both branch off perf.