fix(datagram): preserve quarter stream id

[HeikoBornholdt]

h3-datagram encoded the Quarter Stream ID correctly, but then stored a zeroed buffer in EncodedDatagram. As a result, every outgoing HTTP/3 Datagram was serialized with Quarter Stream ID 0.

This breaks datagrams for later request streams on the same HTTP/3 connection, for example when multiple MASQUE CONNECT-UDP sessions share one H3 connection. Those datagrams appear to belong to the first request stream and may be routed to the wrong session, ignored, or rejected.

[Streetblock]

Tested this commit in an isolated merge against current upstream/master.

• the merge is conflict-free
• cargo fmt --all -- --check passes
• cargo test --workspace passes (227 h3 tests, the h3-datagram test, and 9 documentation tests)
• the only warning is the pre-existing direct-path h3-quinn cfg warning tracked separately in #335

I also expanded the round-trip test locally across all QUIC VarInt width boundaries for Quarter Stream IDs (0, 1, 2^6-1, 2^6, 2^14-1, 2^14, 2^30-1, 2^30, and 2^60-1). All cases preserve the original stream ID and payload.

The one-line production fix correctly uses the already encoded header buffer instead of replacing it with zeros. The broader boundary table could be a useful optional test enhancement, but I do not consider it a merge blocker.

[seanmonstar]

Looks good, thanks!