fix(spop): negotiate frame sizes lazily

[fionera]

Stacked on the handshake tests merged in #32.

Summary

• process HAPROXY-HELLO before scheduling later frames and echo its max-frame-size
• inherit HAProxy's advertised limit and reject values above HAProxy's signed-int ceiling
• treat negotiation as a per-connection validation limit, not an allocation request
• grow shared pooled buffers only when an actual inbound frame needs more space
• make the existing ActionWriter grow internally on demand and enforce the negotiated ACK limit

HAProxy advertises tune.bufsize - 4; there is no universal 64 KiB cap. HTTP configurations stay below 256 MiB, while TCP/SPOE configurations can approach 2 GiB, so lazy allocation is still required.

Memory behavior

Escape analysis confirms the frame and backing slices are heap allocated. sync.Pool retains grown buffers for reuse; the previous comment about a maximum stack slice size was incorrect. A large advertised value alone does not allocate a large buffer.

Validation

• go test ./...
• go test -race ./...
• GOARCH=386 go test ./...
• GOTOOLCHAIN=go1.21.13 go test ./...
• go test -tags=e2e ./spop
• go vet ./...
• staticcheck ./...

Fixes #14