perf(ext/node): convert more polyfills to lazy-loaded JS

[bartlomieju]

Summary

Continues the lazy-loading conversion started in #33876 and #33881.

• Convert 6 node:* modules from eagerly-loaded ESM to lazy-loaded IIFE scripts with thin ESM wrappers: node:wasi, node:trace_events, node:punycode, node:domain, node:perf_hooks, node:vm
• Update import sites in 01_require.js, http2.ts, and repl.ts to use core.loadExtScript() instead of ESM imports
• Add trace_events to DENO_SUPPORTED_BUILTIN_NODE_MODULES (was previously missing from the builtin list)

Test plan

• ./x test-spec unknown_builtin_node_module — updated expectations, passes
• ./x test-spec node_builtin — 11 tests pass
• ./x test-spec domain — domain timer tests pass
• cargo test unit_node::vm — passes
• cargo test unit_node::domain — passes
• cargo test unit_node::perf_hooks — passes
• cargo test unit_node::punycode — passes
• cargo test unit_node::http2 — passes (import change)
• cargo test unit_node::repl — passes (import change)
• Smoke tested all 6 modules with ESM import and CJS require()

[fibibot]

Continuation of the lazy-load lineage. Substance follows the established pattern from #33881 / #33876 — looks right for the identity contract this time around.

What I verified

The vm.js IIFE return shape (representative of all 6 conversions):

return {
  default: { Module, Script, ..., compileFunction, measureMemory },
  Module, Script, ..., compileFunction, measureMemory,  // top-level for named imports
};

vm_esm.js wrapper:

const mod = core.loadExtScript("ext:deno_node/vm.js");
export const { Module, Script, ... } = mod;
export default mod.default;

01_require.js consumer:

const vm = core.loadExtScript("ext:deno_node/vm.js").default;

So imported.default (ESM) and require("vm") (CJS) both resolve to mod.default — same object, identity preserved. Should pass process.getBuiltinModule("vm") === imported.default (the test that broke #33875). ✓

Trace events fix: trace_events was missing from DENO_SUPPORTED_BUILTIN_NODE_MODULES — adding it cleans up the spec test for unknown_builtin_node_module (5 lines removed from check.out). Worth a sanity check that tools/lint.js won't flag this as an unrelated change in the same PR, but it's a minor cleanup.

CI

56 success, 10 pending, 0 fail. All the substantive shards (build/lint/test_specs/test_node_compat) haven't reported yet. Will re-confirm once they land. The 6 modules being converted (wasi, trace_events, punycode, domain, perf_hooks, vm) are exactly what test-process-get-builtin.mjs tests — that test will be the load-bearing signal for whether identity contract is preserved.

Not APPROVE-ing yet per the CI-must-be-green rule; substance looks right but the test that broke #33875 is the one I want to see green.

[fibibot]

Substance is correct

The test-process-get-builtin.mjs family (which broke #33875 with the same conversion shape) passes on all 4 platforms × debug + release. All 14 test node_compat (1/3, 2/3) shards green — confirming the identity contract is preserved through this lazy-load conversion. ✓

One unrelated-Windows fail

test node_compat (3/3) debug windows-aarch64 and windows-x86_64 both fail with one test:

---- node_compat::parallel::test-child-process-send-returns-boolean.js ----
error: Uncaught TypeError: This handle type cannot be sent
  const rv1 = s.send('one', handle, assert.ifError);

This test was enabled in #33869 (which I approved earlier today) and the failure surface is child.send(msg, handle) IPC-handle passing on Windows — completely unrelated to this PR's scope (lazy-loading wasi / trace_events / punycode / domain / perf_hooks / vm). Looking at the established pattern in tests/node_compat/config.jsonc, similar IPC-handle tests already have "windows": false (e.g., test-child-process-send-keep-open.js with reason "net.Socket IPC handle passing uses SCM_RIGHTS, not yet implemented on Windows"). The new boolean-return test should likely have the same gate.

This fail probably blocks merge of the boolean-return test's PR (#33869, already merged) more than it blocks this conversion PR. Worth a 1-line follow-up to add the Windows skip.

CI

131 success, 3 fail (1 = ci status rollup, 2 = the same Windows test), 2 skipping. All non-Windows shards green; Windows-specific test is the only blocker.

Not APPROVE-ing per the strict CI-must-be-green rule, but substantively this PR is APPROVE-quality — the conversion is correct, identity preserved, and the failing test is provably unrelated. If the maintainer adds "windows": false to the boolean-return test (either in this PR's diff or a follow-up that lands first), I can clear with APPROVE.