Add an egress CIDR allowlist to the SSRF guard (closes #204)

The SSRF blocklist had no escape hatch, so the thing webhooker is
mostly for — taking a public webhook and forwarding it to something
on your own network — could not be configured at all. Every private
address, Docker sibling and loopback service was permanently
unreachable as a delivery destination.

ALLOWED_EGRESS_CIDRS (default empty) names blocks that delivery
targets may reach despite the default blocklist. It is an allowlist
and only ever adds destinations: there is no boolean, and no value
disables SSRF protection wholesale. Empty, the guard behaves exactly
as before.

A fixed set of addresses is refused before the allowlist is
consulted, so no supplied CIDR opens one — not the exact address,
not a supernet, not 0.0.0.0/0 or ::/0. It is the two link-local
blocks (169.254.0.0/16, fe80::/10) plus host routes for the cloud
metadata endpoints that sit outside them: AWS's IPv6 IMDS at
fd00:ec2::254, which lives in ordinary ULA space, and Alibaba's
100.100.100.200, which lives in CGNAT. Allowlisting fd00::/8 or
100.64.0.0/10 (Tailscale's range) is an ordinary thing for an
operator to do and must not reopen instance-credential theft. The
IPv4-compatible (::a9fe:a9fe) and NAT64 (64:ff9b::a9fe:a9fe)
spellings of 169.254.169.254 are listed too, because To4() does not
normalise them into the link-local block the way it does the
IPv4-mapped form. Reaching any of these is credential theft rather
than delivery to an internal service.

The policy now lives in one function, Guard.checkIP, which both
target-creation validation and the delivery dialer call. The two
paths previously decided separately, which is how they came to
disagree about a destination. The guard is built once from config
and injected via fx into both the handlers and the delivery engine,
so there is a single instance and a single answer.

A set-but-unparseable value aborts startup naming the variable,
reusing the existing envPrefixList parser. A non-empty list is
logged at startup with the blocks spelled out, not counted, so the
hole is visible in the log of any deployment that has one.

Tests: an allowlisted loopback CIDR both validates and delivers to a
live server (and the same URL still fails without the allowlist); a
private address outside the listed block stays refused on both
paths; every unconditionally blocked address stays refused on both
paths under an allowlist that covers it, and the set itself is
pinned entry by entry; public addresses are unaffected either way;
and config coverage for parsing, startup abort, and the warning's
contents.

internal/delivery/client_ssrf_test.go

@@ -18,8 +18,9 @@ func newSSRFTestEngine() *delivery.Engine {
 	log := slog.New(slog.DiscardHandler)
 
 	client := &http.Client{
-		Timeout:   30 * time.Second,
-		Transport: delivery.NewSSRFSafeTransport(),
+		Timeout: 30 * time.Second,
+		Transport: delivery.NewTestGuard().
+			NewSSRFSafeTransport(),
 	}
 
 	return delivery.NewTestEngine(log, client, 1)
@@ -36,8 +37,8 @@ func TestClientForConfig_TimeoutKeepsSSRFGuard(t *testing.T) {
 	engine := newSSRFTestEngine()
 
 	blocked := []string{
-		"http://127.0.0.1/hook",
-		"http://169.254.169.254/latest/meta-data/",
+		loopbackHookURL,
+		metadataURL,
 		"http://[fe80::1]/hook",
 	}