Question for using std rust channel instead of tokio's channel in async_policy of reinforcement learning

[seongs1024]

Feature description

burn has already had the dependency, tokio, which is efficient for task scheduling.

burn/Cargo.toml

Lines 110 to 112 in 5923b1e

	tokio = { version = "1.49.0", features = ["rt", "macros"] }
	tokio-tungstenite = "0.28"
	tokio-util = "0.7"

But async_policy use std rust channel.

burn/crates/burn-rl/src/policy/async_policy.rs

Lines 1 to 8 in 5923b1e

	use std::{
	sync::{
	Arc,
	atomic::{AtomicUsize, Ordering},
	mpsc::{self, Sender},
	},
	thread::spawn,
	};

I guess, reasons for this are:

1. To avoid context switching overhead
2. Not necessary because messages are processed in a dedicated thread
3. To avoid function coloring

If so, crossbeam may be a candidate for better performance.

Feature motivation

• tokio will be suitable for async non-blocking operations

(Optional) Suggest a Solution

[seongs1024]

crossbeam channel is merged into std mpsc channel