Use jemalloc as global allocator

Production lightning nodes were getting OOM-killed during boot.
The bun process running the NAPI-RS native addon hit 857MB RSS
in a 1GB container, leaving no headroom for channel monitor
deserialization at startup.

glibc's malloc fragments heavily under Rust's allocation
patterns (many small, short-lived allocations interleaved with
long-lived ones), inflating RSS 30-50% beyond actual usage.
jemalloc's arena-based allocator and thread-local caching avoid
this fragmentation.

disable_initial_exec_tls is required because NAPI-RS addons are
loaded via dlopen at runtime. Without it, jemalloc's TLS usage
exhausts the static TLS block allocated at program start, since
dlopen'd libraries get a smaller TLS allocation than statically
linked ones.

Author: amackillop

Cargo.toml

@@ -8,6 +8,8 @@ version      = "0.1.0"
 crate-type = ["cdylib"]
 
 [dependencies]
+tikv-jemallocator = { version = "0.6", features = ["disable_initial_exec_tls"] }
+
 bitcoin-payment-instructions = { version = "0.5.0", default-features = false, features = [
   "http",
 ] }

src/lib.rs

@@ -1,5 +1,11 @@
 #![deny(clippy::all)]
 
+/// Use jemalloc to reduce memory fragmentation during channel monitor
+/// deserialization and node boot. glibc malloc fragments heavily with
+/// Rust's alloc patterns
+#[global_allocator]
+static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
+
 use std::{
   collections::{HashMap, HashSet},
   convert::TryFrom,