feat: add optional jemalloc global allocator (#402)

[mwiewior]

Summary

Adds an opt-in jemalloc Cargo feature (off by default) that installs
tikv-jemallocator as the global
allocator. Addresses #402.

• Off by default — enable with maturin build --features jemalloc (Linux/macOS).
• Gated to non-MSVC targets (cfg(not(target_env = "msvc"))), so Windows keeps the
  system allocator (tikv-jemallocator doesn't support MSVC).
• Default builds are byte-identical to before — no allocator is compiled in unless
  the feature is enabled.

#[cfg(all(feature = "jemalloc", not(target_env = "msvc")))]
#[global_allocator]
static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;

Why jemalloc (and not mimalloc)

I benchmarked default vs mimalloc vs jemalloc on the single-thread interval-operation
suite (databio, 8 operations, POLARS_MAX_THREADS=1, -C target-cpu=native, macOS
arm64). Two reasons jemalloc wins:

1. jemalloc is faster and leaner. Runtime geomean −12.3% vs default, with
   equal-or-lower peak RSS. mimalloc was faster too (−9.3%) but cost +36–40% peak RSS
   on reduce ops (merge/complement) because it retains freed segments.
2. mimalloc v3 (the crate default) segfaults at import on macOS arm64 — a load-time
   EXC_BAD_ACCESS in mi_bbitmap_try_find_and_clear. It only works if pinned to the
   v2 feature. jemalloc has no such issue.

Runtime, vs default (geomean of per-case ratios)

operation	jemalloc Δ%
merge	−23.5%
count_overlaps	−17.6%
coverage	−16.1%
cluster	−12.8%
complement	−12.0%
nearest	−7.0%
overlap	−4.1%
subtract	−2.3%
overall	−12.3%

Peak memory (whole-run RSS, identical workload)

allocator	peak	Δ vs default
default (system)	14.95 GiB	baseline
jemalloc	14.19 GiB	−5.0%
mimalloc v2	21.19 GiB	+41.8%

Numbers are macOS-laptop, single-thread, directional — macOS libmalloc is slow, so
the win overstates a Linux release build. Full data (per-op runtime + per-op peak/avg
RSS) lives in the polars-bio-bench repo (results/alloc-impact-summary.md).

Verification

• RUSTFLAGS="-Dwarnings" cargo check (default) and --features jemalloc both clean.
• Built the --features jemalloc wheel and confirmed import polars_bio + an overlap
  run succeed on macOS arm64.

Follow-ups (not in this PR)

• Since the feature is off by default, the published PyPI wheels do not include it.
  To ship jemalloc to end users, add --features jemalloc to the maturin args in the
  Linux/macOS jobs of .github/workflows/publish_to_pypi.yml (leave Windows untouched).
• A Linux, ideally multi-threaded, confirmation run is worth doing before flipping it
  on in wheels — multithreaded allocation is where jemalloc typically helps most.

🤖 Generated with Claude Code