fix(q): harden response handling and encode dict results

[belowzeroff]

How it looks from the user's side

A q client asks a Rayforce Q-wire server for a native dictionary result:

h "(dict ['a 'b] [1 2])"

Before -- the server failed to serialize the native RAY_DICT, so the user saw a Q error instead of the dictionary:

test/rfl/server/types.rfl:45: LHS error: error: q: cann  -- src: (.q.send h "(dict ['a 'b] [1 2])")

After -- the same expression round-trips as a Q dictionary:

T["dict"; (`a`b!1 2)~h"(dict ['a 'b] [1 2])"]
real-q interop ok (30 assertions)

A client asking real q for the identity value also now gets a Rayforce null instead of a decode error:

(nil? (.q.send h "::")) -- true

What changed

Fixes several Q IPC response edge cases:

• Rejects non-response frames in blocking q_exchange, so pushed or malformed frames are not mistaken for sync replies.
• Encodes native RAY_DICT values as Q dictionaries on the wire.
• Decodes Q identity (::, wire type 101 primitive 0) as Rayforce null.
• Suppresses SIGPIPE / MSG_NOSIGNAL on send, so closed peers fail as ordinary send errors instead of terminating the process.
• Uses a release helper for ray_error values in decode/server cleanup paths.

Tests

make test

Covered:

• codec selftest
• exchange selftest
• Rayforce server round-trip tests
• malformed Q frame handling
• real-q interop
• client tests against real q
• event-loop client and push tests