perf: avoid atomic refcount bumping on the dispatch hot path by replacing load_full() with load()

[shivampipalwa]

Inside dispatcher logic in ConnectionManager, I noticed an opportunity to eliminate unnecessary atomic operations on the hot path.

Currently, dispatch_account_update, dispatch_transaction_update, and dispatch_slot_update iterate over self.connections using slot.load_full():

code

for (idx, slot) in self.connections.iter().enumerate() {
    let Some(entry) = slot.load_full() else { 
        continue;
    };
    // ...
}

Bottleneck:
load_full() performs an atomic increment/decrement on the underlying Arc's reference count. On a high-throughput Solana stream (e.g., 50k+ updates/sec), doing this for every slot on every update causes significant cache-line bouncing and memory bus contention.

Proposed Solution:
Switch the hot-path iteration to use slot.load() instead, which returns a Guard.

for (idx, slot) in self.connections.iter().enumerate() {
    let guard = slot.load();
    let Some(entry) = guard.as_deref() else {
        continue;
    };
    // ...
}

Because the Guard is scoped to the loop iteration and dropped immediately, it utilizes arc-swap's thread-local hazard pointers (staying well under the cheap proxy limit) and avoids the atomic reference count bump. We can pass &ManagedConnection to the helper functions down the stack.

I'd love to contribute this fix. Let me know if you're open to a PR!