build: parallelize native compile with xargs -P, not batch-and-wait

[dfa1]

Problem

The compile loop backgrounded up to $JOBS zig cc invocations, then hit a
bare wait every $JOBS iterations before starting the next batch:

i=0
for src in $SRCS; do
    zig cc -target "$ZIG_TARGET" $CFLAGS -c "$src" -o "$out" &
    i=$((i + 1))
    [ $((i % JOBS)) -eq 0 ] && wait
done
wait

Two problems with this:

1. Batch barrier, not a work queue. Every JOBS compiles, every core
   idles until the slowest TU in that batch finishes before the next batch
   starts — no work-stealing across the boundary.
2. Failures are swallowed. A backgrounded zig cc failing under & is
   invisible to set -e — a bare wait returns success regardless of the
   child's exit status. A broken compile silently produces no .o and only
   surfaces later as a cryptic link error (missing symbols) or, worse, a
   link that "succeeds" against a stale .o left over from a previous run.

Fix

Replace the loop with xargs -P "$JOBS", which keeps all $JOBS slots
continuously saturated (no batch boundary) and returns non-zero if any
invocation fails, which set -e now catches immediately.

compile_one() {
    zig cc -target "$ZIG_TARGET" $CFLAGS -c "$1" -o "$WORK/$(basename "$1").o"
}
export -f compile_one
export ZIG_TARGET CFLAGS WORK

printf '%s\n' $SRCS | xargs -P "$JOBS" -I{} bash -c 'compile_one "$@"' _ {}

-I{} + bash -c '... "$@"' _ {} is the standard idiom for passing an
xargs-substituted value into bash -c as $1 rather than splicing it
unquoted into the command string.

Testing

Verified locally with zig cc (osx-aarch64 target):

• Clean build still produces a working libzstd.dylib.
• Injected a syntax error into xxhash.c and reran — previously this kind
  of failure was invisible; now the script exits 1 immediately with the
  compiler's error printed, before ever reaching the link step.

Addresses point 1 of the review feedback on the native build script.