ltc: Send() socket deadline + desync-safe teardown (parity with dash #781)

[frstrtr]

P1 — live LTC-DOGE prod reference: RPC socket deadline

Propagates the DASH #781 (9ac8b3dd) socket-deadline class to LTC, the P1 lane
in the cross-coin io/RPC reconciliation audit. LTC already has the class-A
thread_pool decouple and the class-B m_rpc_mutex; the remaining gap is the
class-C deadline.

Problem

ltc::coin::NodeRPC::Send() has no deadline — a wedged-but-connected
litecoind/dogecoind (socket open, no bytes) hangs the blocking http::read
forever → permanent refresh-wedge + SIGINT-hang on join, on the live merged-
mining node.

Change — port DASH's mechanism verbatim

• apply_socket_timeouts() — forces the socket back to BLOCKING mode
  (async_connect leaves asio's non_blocking flag set, under which SO_*TIMEO
  is a no-op) then sets kernel SO_RCVTIMEO/SO_SNDTIMEO =
  RPC_IO_TIMEOUT_SECONDS (12). Called after every successful (re)connect (the
  async_connect handler before check(), and sync_reconnect()). beast's
  synchronous read_some/write_some do not honour
  tcp_stream::expires_after (async-only), so the kernel timeout is the robust
  bound. #ifndef _WIN32 guarded (Linux release target; Windows keeps the
  pre-parity no-deadline behaviour).
• close_stream() — desync-safe teardown on the final-attempt Send()
  exits (both write-fail and read-fail). This is the critical half — the
  deadline alone reintroduces dash: decouple blocking coin-RPC + heavy work off the stratum io_context (io-thread-decouple) #781's item-2b off-by-one desync.

Why the desync applies to LTC (verified, not assumed)

LTC's NodeRPC shares DASH's shape: const std::string ID = "curltest" (a
constant request id) driving jsonrpccxx::JsonRpcClient, which never validates
response ids. A timed-out-but-still-open stream (request written, response
unread) reused by the next Send() would read this request's late response
→ permanent off-by-one: a GBT parses a getbestblockhash hash-string → zeroed
work → a stuck "honest set-gap" outage until the connection churns.
close_stream() tears the socket down so the next Send() write-fails into a
clean sync_reconnect(). LTC already had m_rpc_mutex (class B) — no mutex
work needed.

Consensus-neutral

Transport/liveness only — zero wire-byte change; strict added-code diff-grep
for share/target/payout/vardiff/reward/coinbase/merkle/nbits/difficulty is
empty. Satisfies the pre-v36 transition rule.

Scope note — LTC class-D (WebServer) is a separate follow-on

LTC's miner-facing server is core::WebServer, not core::StratumServer,
so the #781 core reap knobs do NOT reach LTC. Its zombie-session class needs a
WebServer-level keepalive + idle-reap fix — queued, out of scope for this PR
(class-C deadline only).

Build/test

c2pool-ltc builds clean with the change compiled in; LTC-lane regression
(test_template_builder 35/35, test_doge_chain 37/37) green. NodeRPC is not
unit-testable without a live daemon socket.

Draft — not for merge/deploy; LTC-DOGE prod is an operator tap. Gets an
LTC-parity review first.