Use jemalloc as global allocator on unix

[briangmilnes]

Summary

• On Linux, switching from glibc malloc to jemalloc saves avg 12%, peak 14% RSS
  and avg 9%, peak 14% wall time across 6 open-source Verus projects, with zero
  code changes
• cfg(unix) gated — no effect on Windows builds
• unprefixed_malloc_on_supported_platforms feature ensures jemalloc replaces
  C-level malloc/free globally, not just Rust's allocator (required because
  librustc_driver.so uses C malloc via LLVM)

Benchmark results

#	Project	Verified	glibc MB	jemal MB	RSS Δ	glibc s	jemal s	Time Δ
1	human-eval	292	614	537	-13%	5.4s	5.1s	-4%
2	ironkv	319	783	706	-10%	5.4s	5.1s	-5%
3	node-replication	254	957	854	-11%	6.5s	5.5s	-14%
4	vest	496	648	565	-13%	5.7s	5.2s	-9%
5	memory-allocator	731	1,629	1,416	-13%	38.7s	35.2s	-9%
6	APAS-VERUS	4,265	5,822	5,143	-12%	59.2s	53.2s	-10%
	Peak				-14%			-14%
	Average				-12%			-9%

macOS is not yet benchmarked. jemalloc is not available on Windows.

Full analysis: https://github.com/briangmilnes/verus-heaptrack

Test plan

• vargo build --release succeeds, vstd rebuilds clean (1490 verified)
• vargo test --package rust_verify_test — 3928 passed, 0 failed
• Verified 6 open-source Verus projects with jemalloc-built binary

[briangmilnes]

Double checked the build and the benchmarks looking for any crashes, and it's clean.