perf(core): drop the per-op .catch wrapper promise

[bartlomieju]

Every pending async op currently allocates two promises: the op promise created in setPromise, plus a second, derived promise from PromisePrototypeCatch(promise, __opRejectHandler) whose only purpose is to recapture the error stack (stripping internal event-loop frames) on the rare rejection path. This doubles the promise objects the GC sweeps per in-flight op and registers promise-reaction machinery on every op.

This PR returns the raw op promise (still carrying promiseIdSymbol, so refOp/unrefOp keep working) and drops the recapture.

Impact

• ~6.5% faster across all tick-path ops on the async op microbench (libs/core/benches/ops/async.rs), measured back-to-back:

  bench (1000 ops/iter)	before	after	Δ
  op_async_void_deferred	429,487	401,318	−6.6%
  op_async_void_deferred_nofast	433,513	403,950	−6.8%
  op_async_void_lazy	425,649	394,058	−7.4%
  op_async_yield_deferred	423,214	396,355	−6.3%
  op_async_yield_lazy_nofast	535,074	498,536	−6.8%

• One fewer promise + reaction allocated per in-flight async op → less GC pressure on op-heavy servers.

Stack traces

The recapture only ever affected a bare op stub awaited directly (no JS wrapper). Every real Deno async API (Deno.readFile, fetch, …) is an async function that awaits the op, so V8's async stack traces already include the call site without the recapture. Verified caught, uncaught, and dynamic-import error stacks via deno run are byte-identical to before.

The one raw-runtime test that exercised the bare-op pattern (test_dynamic_import_module_error_stack) is updated: it now asserts on the await import(...) site in main.js rather than the innermost bare-op frame. The user-facing #20034 guarantee stays covered by the CLI specs error_023_stack_async and fetch_async_error_stack (both pass).

Tests

• cargo test -p deno_core (441/441 relevant, incl. updated dynamic-import stack test)
• spec: error_02* stack tests, fetch_async_error_stack, promise/rejection tests — all pass