perf: reduce memory usage of Command.output()

[bartlomieju]

Summary

Fixes RSS memory growth in Deno.Command.output() when called in a loop. Over 30k iterations, RSS grew from ~70MB to ~2GB while JS heap stayed flat at ~18MB.

The Command.output() path was creating ReadableStream wrappers (via readableStreamForRidUnrefable + readableStreamCollectIntoUint8Array) for stdout/stderr just to immediately collect all their data. The stream/controller machinery has significant overhead that causes RSS growth in tight loops.

The fix bypasses both ChildProcess and ReadableStream for the Command.output() hot path. spawnInner now:

• Calls op_spawn_child directly instead of creating a ChildProcess instance
• Reads stdout/stderr via core.read in a simple loop (readAllRid)
• Avoids the overhead of stream creation, BYOB readers, and controller state

Before: 30k iterations -> RSS ~2GB
After: 30k iterations -> RSS ~215MB (stable)

Note: ChildProcess.output() (via .spawn().output()) is unchanged and still uses streams.

Test plan

Repro script from the issue:

for (let i = 1; i <= 30000; i++) {
  const command = new Deno.Command('echo', { args: ['hello'] });
  const result = await command.output();
  if (i % 5000 === 0) {
    const mem = Deno.memoryUsage();
    console.log(`${i}: RSS=${(mem.rss / 1024 / 1024).toFixed(1)}MB heap=${(mem.heapUsed / 1024 / 1024).toFixed(1)}MB`);
  }
}

Closes #33334
Closes #24674

🤖 Generated with Claude Code

[miracatbot]

LGTM.

I took a close look at the new Command.output() fast path in ext/process/40_process.js.

The change makes sense to me:

• it bypasses the ReadableStream wrapper machinery for the hot Command.output() case,
• reads stdout/stderr directly with core.read() in a simple loop,
• still waits for process completion in parallel,
• and preserves the existing user-facing behavior around piped vs non-piped stdio.

I didn't spot a concrete correctness issue in the new resource-lifetime or output-collection logic, and the motivation / repro is compelling.

[nathanwhit]

I think the PR description isn't quite right but the change is a perf improvement and seems right, so LGTM