Only panic when sending large packets if in debug mode

[caspark]

What it says on the tin.

I also moved the packet size checking to the UDP socket layer, since technically the thresholds will be different for different network transports.

---

• This isn't a fatal error so I don't think crashing in release mode is appropriate.
• This unblocks RFC: feat: Input type must be Default+Serde but not POD #82 - without a hard cap on input sizes, I think it's sane to lift the POD restriction on input types.

Also, the original panic-always behavior is super annoying when you tell ggrs your input type is "arbitrary buffer of bytes" and rely on the delta (XOR + run length) encoding to shrink it down, because if your "real" input type has e.g. optional or variable sized data in it (like a vec of positions, or similar), then the delta encoding won't shrink the message below the packet size cut-off - and then on high latency connections with long send queues ggrs will crash.

(That issue could also be improved somewhat by improving the delta encoding: currently delta encoding encodes xors each input against the last confirmed input, but I think it should be possible to delta encode against the last confirmed input N for the input N+1, then for N+2 encode vs N+1, N+3 encodes vs N+2, etc - and ideally factor in custom input prediction too. But that's much more involved than this PR!)

In my personal fork of ggrs, I have actually added logging and I think that's the better approach, but that's a relatively invasive change - so I stuck with the conservative approach here.

[gschup]

Good change!