We currently pass a single large read buffer to the OS across all platforms:

Linux, Android and Windows can leverage GRO (and URO) to read multiple datagrams into the buffer.

Apple platforms do not support GRO. But instead, we can fall-back to the older recvmmsg style syscalls.

@larseggert added recvmmsg style reads on apple to quinn-udp: quinn-rs/quinn#1993

In order to make use of this optimization on apple platforms, we should be passing multiple smaller read buffers to the apple OS, instead of one large buffer (see code above).

I suggest the following:

1. Introduce a struct ReceiveBuffer(Vec<u8>) wrapper type with a new function using RECV_BUF_SIZE.
2. Have the user pass a long-lived instance of ReceiveBuffer to recv_inner.
3. Within recv_inner we can assert the size of receive_buffer.0, given that a user can construct it via new only. On Linux, Android, Windows, ... proceed as before, passing the single large buffer to the OS. On Apple, partition the buffer and pass multiple slices of it to the OS.
4. On apple platforms, enable the fast_apple quinn-udp feature.

Alternatively on Apple platforms, we could change ReceiveBuffer(Vec<u8>) to ReceiveBuffer(Vec<Vec<u8>>).

Overarching tracking issue: #1693