fix: enable unprefixed jemalloc to override musl malloc for SQLite C code

[ahmedhesham6]

Description

Fix autopilot SIGSEGV crash-loop on Linux musl by making jemalloc actually override the system malloc for all code in the process, including SQLite's embedded C.

Root Cause

tikv-jemallocator without unprefixed_malloc_on_supported_platforms only provides prefixed symbols (_rjem_je_malloc). The #[global_allocator] only routes Rust allocations through jemalloc. SQLite's C code (compiled via libsql-ffi) calls malloc() directly — which on a statically-linked musl binary is musl's aggressive allocator that munmaps freed pages.

Confirmed via GDB core dump on i-0773f92db43188a0c (eu-north-1):

#0  sqlite3Close+25: movzbl 0x71(%rdi),%eax   ← reading unmapped page
#1  drop_in_place<libsql::local::connection::Connection>
#2  Arc::drop_slow()
#3  run_scheduler::{{closure}}

And via nm: malloc was musl's weak (W) symbol, not jemalloc.

Changes Made

• cli/Cargo.toml: Add unprefixed_malloc_on_supported_platforms feature to tikv-jemallocator so jemalloc provides the real malloc/free symbols
• cli/src/main.rs: Document why the feature is critical
• build-and-release.yml: Re-enable aarch64-unknown-linux-musl builds (same root cause was blocking them); fix test step to pass --features jemalloc on musl
• musl-jemalloc.yml (new): CI workflow that verifies via nm that malloc is not musl's weak symbol
• test-jemalloc-fix.yml (new, temporary): Full build+test+verify for both arches — delete after confirming

Testing

• cargo check --features jemalloc passes
• cargo test --workspace --lib passes
• cargo test --workspace --lib --features jemalloc passes
• cargo fmt --check clean
• cargo clippy --all-targets clean
• Temp CI validates malloc symbol override on x86_64 + aarch64

Breaking Changes

None