fix(prover): print per-table instruments report on the continuation path

[MauroToscano]

Bug

The instruments cargo feature prints a rich per-table-kind timing report
(merges split tables by kind, e.g. MEMW[0..4] -> MEMW x5, aggregates
count / rows / duration / sub-ops, sorted by cost). That report was only ever
emitted on the monolithic prove() path, from the
#[cfg(feature = "instruments")] block in prover/src/lib.rs.

The continuation path never went through that block: prove_epoch and
prove_global in prover/src/continuation.rs call Prover::multi_prove
directly. multi_prove does instruments::store(MultiProveTiming { .. }) for
every call, so the per-table data was already being computed and stored — but
nothing on the continuation path ever printed it. Result:
prove --continuations with --features instruments produced no per-table
report at all.

Fix

Feature-gated only (zero cost, and byte-for-byte identical compile, when the
feature is off):

• prover/src/instruments.rs: factor the mp-derived per-table block out of
  print_report into a shared print_multi_prove(mp, total) helper, and add
  print_epoch_report(label) which drains the timing stored by the preceding
  multi_prove via stark::instruments::take() and formats it with that same
  helper (so both paths format from a single source of truth). Also inlined a
  literal into one eprintln! format string to silence a pre-existing
  clippy::print_literal that only trips under --features instruments;
  rendered output is unchanged.
• prover/src/continuation.rs: after each Prover::multi_prove, call
  print_epoch_report, labeled EPOCH N for each epoch (from the 1-based
  EpochStart.label) and GLOBAL for the cross-epoch memory proof.

The monolithic path's output is unchanged (same statements, same order — the
mp block was moved into a function and called with the same mp/total). No
proving logic changed; this is purely gated reporting.

Note on GLOBAL: its per-epoch total is derived from the sum of the proving
phases, so for a trivially small global proof (a couple of tiny tables that run
in parallel) individual table percentages can exceed 100% — an honest artifact
of parallel per-table timing against a tiny sequential-phase denominator (the
monolithic formatter has the same parallel overlap, just diluted by a larger
wall-clock total). The EPOCH N reports, which carry the real work, read
normally.

Verification

Built and ran a continuation prove that exercises prove_epoch (x2) +
prove_global:

cargo build --release -p cli --features instruments
./target/release/cli prove executor/program_artifacts/asm/fib_iterative_500k.elf \
    --output /tmp/cont.proof --continuations --epoch-size-log2 18

Per-epoch reports now print. Sample (trimmed):

=== EPOCH 0 ===
  Phase                                    Wall      %
  --------------------------------------------------------
  Pre-pass (domains/twiddles)             0.03s    1.0%
  Round 1                                 1.30s   38.8%
    Main trace commits                    0.59s          17.7%
    ...
  Rounds 2-4                              2.01s   60.1%
      BITWISE              1.0M           0.82s          24.5%
      MEMW_R               524K           0.47s          13.9%
      EQ                    66K           0.32s           9.7%
      CPU                  262K           0.32s           9.6%
      ...
      (16 others)                         0.67s          20.0%
      -- sub-operation totals (all tables) --
      R4  queries & openings              1.09s          32.7%
      R3  OOD evaluation                  0.66s          19.7%
      ...
  Total FFT                               0.79s   23.6%
  Total Merkle                            0.86s   25.7%
  --------------------------------------------------------
  TOTAL                                   3.34s

=== EPOCH 1 ===
  ... (same layout) ...

=== GLOBAL ===
  ... (same layout) ...

• Feature ON: reports print per epoch (EPOCH 0, EPOCH 1, GLOBAL).
• Feature OFF (cargo build --release -p cli): zero instruments lines in the
  continuation prove output; compiles clean (all new code is
  #[cfg(feature = "instruments")]).
• cargo test --release -p lambda-vm-prover --lib: 521 passed, 0 failed.
• clippy --features instruments (which CI's make lint does not exercise)
  is clean on the changed crates; all three make lint clippy passes exit 0;
  cargo fmt --check is clean on the changed crates.