perf: pre-allocate HashMap and remove redundant intermediate Vec in multiproof.rs

[KENILSHAHH]

Summary

• Pre-allocate FxHashMap in group_prover_queries with with_capacity_and_hasher(prover_queries.len(), ...) to avoid repeated rehashing as entries are added
• Remove intermediate collect into Vec + manual par_chunks(len / num_threads) splitting in MultiPoint::open, replacing with direct into_par_iter().map() which
  eliminates one Vec allocation and lets rayon's work-stealing handle load balancing

Context

Both changes are in ipa-multipoint/src/multiproof.rs, in the MultiPoint::open proof generation hot path.

The HashMap was starting at default capacity and resizing multiple times as queries were inserted. Since we know the number of queries upfront, we pre-allocate.

The old parallelization pattern collected a HashMap into a Vec just to call par_chunks() with a manually computed chunk size. This added an unnecessary allocation and
rigid equal-sized splitting. Rayon's into_par_iter() on HashMap already provides parallel iteration with work-stealing, which handles uneven workloads better than fixed
chunking.