broadcast: parked-chunk buffering + poll-driven TTL cleanup

[ch4r10t33r]

Closes #63.

Ports the channel parked-chunk buffer and GC from ethp2p broadcast/channel.go (handleChunk / handleSessionOpen / cleanup). ChannelRs previously had no parked map or TTL handling — chunks arriving before their SESS open had nowhere to go, and stale sessions were never reaped.

What it does

• parkChunk buffers a chunk that arrived before its session exists; per-message cap max_parked_chunks = 32 (new chunk dropped → .dropped_full when full, matching Go cancelling the stream).
• attachRelaySession replays and frees any parked chunks into the now-open session (Go handleSessionOpen flush).
• disposeSession also drops the message's parked chunks.
• cleanup(now_ms) disposes sessions older than active_session_ttl_ms and drops parked buckets older than pending_chunk_ttl_ms.
• Adds the constants channel_inbox_cap (1024), max_parked_chunks (32), active_session_ttl_ms (5m), pending_chunk_ttl_ms (10s), cleanup_interval_ms (30s), routing_tick_interval_ms (25).

Adaptations

• Go parks the QUIC stream and lets flow control back-pressure the sender; the Zig ingest path already holds the read bytes, so ParkedChunk owns a copy of the peer id + data.
• cleanup is poll-driven with an explicit now_ms (no ambient clock / ticker goroutine). SessionRs.created_at_ms carries the creation time — 0 means unset and is skipped, set by the driver that owns the clock. (Go's cleanup currently drops parked chunks unconditionally with a TODO to use pendingChunkTTL; this implements that intent.)

Tests (leak-checked)

• replay-on-attach (verified via a recording observer's chunk_rcvd),
• cap enforcement (33rd park → .dropped_full),
• cleanup drops stale parked chunks + disposes expired sessions (and leaves fresh ones).

Verified on stock Zig 0.16.0: zig fmt --check, zig build test, zig build test-broadcast (TSan). README updated.