perf(actions): cache successful runner token verification (#39934)

Every Actions runner RPC (`FetchTask` polling, `UpdateTask`,
`UpdateLog`, ...) goes through the `withRunner` interceptor, which
verifies the runner token with PBKDF2 (`auth_model.HashToken`, ~2 ms).
In a 60s CPU profile of a production instance with many runners, this
accounted for ~49% of all CPU samples.

- Cache successful verifications in a dedicated LRU of up to 10000
runners, keyed by the stored salt and hash and holding only a SHA-256
digest of the token. A hit costs ~100 ns.
- The runner is still loaded from the DB on every RPC and failures are
never cached, so deleted runners and wrong tokens never pass.
- The shared `auth_model.TokenCache` is not used because its default 20
entries miss on every poll once more runners poll in turn.

---------

Co-authored-by: silverwind <me@silverwind.io>

routers/api/actions/runner/interceptor_test.go

@@ -0,0 +1,26 @@
+// Copyright 2026 The Gitea Authors. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+package runner
+
+import (
+	"crypto/sha256"
+	"testing"
+
+	actions_model "gitea.dev/models/actions"
+	auth_model "gitea.dev/models/auth"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestVerifyRunnerToken(t *testing.T) {
+	verifiedRunnerTokens.Purge()
+	runner := &actions_model.ActionRunner{TokenSalt: "salt", TokenHash: auth_model.HashToken("token", "salt")}
+	assert.True(t, verifyRunnerToken(runner, "token"))
+	assert.True(t, verifyRunnerToken(runner, "token"))
+	assert.False(t, verifyRunnerToken(runner, "wrong"))
+
+	other := &actions_model.ActionRunner{TokenSalt: "salt", TokenHash: auth_model.HashToken("other", "salt")}
+	assert.False(t, verifyRunnerToken(other, "token"))
+	assert.Equal(t, [][sha256.Size]byte{sha256.Sum256([]byte("token"))}, verifiedRunnerTokens.Values())
+}