fix: allocate non-gate selectors at trace-active size instead of dyadic size

[johnathan79717]

Summary

• Allocate non-gate selectors (q_m, q_c, q_l, q_r, q_o, q_4) at trace_active_range_size() instead of dyadic_size(), using virtual zeroes for the rest
• Fix get_polynomial_size() to return virtual_size() instead of size(), since the logical polynomial size is the dyadic circuit size
• Fix batch_polynomials in Hypernova to handle polynomials with different backing sizes (non-gate selectors now have smaller backing than other entities like table polynomials)

Benchmark results

cd barretenberg/cpp/build
BB_VERBOSE=1 BB_BENCH=1 ./bin/ultra_honk_bench --benchmark_filter="construct_proof_ultrahonk_1M_gates_dyadic_2_20$" --benchmark_repetitions=1
BB_VERBOSE=1 BB_BENCH=1 ./bin/ultra_honk_bench --benchmark_filter="construct_proof_ultrahonk_1M_gates_dyadic_2_21$" --benchmark_repetitions=1

Circuits ~2000 gates apart straddling the 2^20 boundary:

Metric	Before	After
Peak RSS (2^21 dyadic)	2854 MiB	2375 MiB (-479 MiB, -17%)
Memory gap 2^20 vs 2^21	603 MiB	126 MiB (-79%)
Peak RSS (2^20 dyadic)	2251 MiB	2249 MiB (unchanged)

Test plan

• ultra_honk_tests — 260 passed, 5 skipped
• chonk_tests — 20 passed
• circuit_checker_tests — 80 passed
• hypernova_tests — 9 passed

Resolves AztecProtocol/barretenberg#1625.

[ledwards2225]

@johnathan79717 I'm not sure how readily we can do this but it should also be possible to download only the SRS content that we need. Worth investigating

[johnathan79717]

@ledwards2225 Good idea — the CRS is currently loaded at full dyadic_size() and the Pippenger point table doubles it, so at 2^21 that's ~256 MiB. Loading only what's needed for the active trace could save ~128 MiB at the dyadic boundary. I'll investigate as a follow-up.

[ledwards2225]

LGTM - One conservative suggestion for a bit of extra safety