fix(ext/node): return real OS file descriptors from node:fs APIs

[fraidev]

Closes #6529
Closes #23012
Closes #31441
Closes #31629

Co-authored-by: @sigmaSd

Needs:

• fix(ext/node): handle raw fd instead rid in isatty #31912

Deno's node:fs polyfills previously returned Deno Resource IDs (RIDs) from fs.open(), which are opaque internal indices. Node.js returns real OS file descriptors (integers like 3, 4, 5...), so code relying on fd semantics, including passing fds between worker threads — would break.

This PR implements a JavaScript-side FD-to-RID mapping layer:

• fs.open() / fs.openSync() now return real OS file descriptors
• All fd-consuming APIs (read, write, fstat, fsync, ftruncate, fchmod, fchown, etc.) translate FD → RID before calling internal Deno ops
• fs.close() unregisters the mapping after closing
• Cross-worker fd passing works via dup() — when a worker receives an fd it didn't open, getRid() lazily dups it to create a local resource

About Cross-worker fd passing

When a worker receives an fd via postMessage():

1. The fd integer arrives as a plain number
2. getRid(fd) doesn't find it in the worker's map
3. op_node_dup_fd(fd) creates a dup'd copy owned by the worker
4. The worker can now read/write independently
5. Closing in the worker only closes the dup, the original fd in the parent is unaffected

This matches Node.js behavior where OS fds are process-wide.

[lucacasonato]

About Cross-worker fd passing
[...]
Closing in the worker only closes the dup, the original fd in the parent is unaffected
This matches Node.js behavior where OS fds are process-wide.

these statements seem inconsistent

[lucacasonato]

This is a dangerous op. It allows you to create a Resource for any open fd that Deno has (e.g sqlite DB)

[fraidev]

Indeed, what do you think about separating the internal FDs? A separate set for (SQLite, etc.)

[littledivy]

I think the FD -> resource map should be in Rust so we can verify the fd exists in the map before calling dup() on it

[lucacasonato]

FDs are assigned by the kernel - we can't have a different set for internal tasks.

[fraidev]

About Cross-worker fd passing
[...]
Closing in the worker only closes the dup, the original fd in the parent is unaffected
This matches Node.js behavior where OS fds are process-wide.

these statements seem inconsistent

Why? They are valid for all process (read and write), but their lifetimes are duplicates.