fix(derp): DERP TLS session resumption, longer connect timeout, upgrade backoff

[fudio101]

Summary

Adapted from cplewes/microlink@b9636816 (issue #25). Not a literal cherry-pick: git apply --check --3way conflicted only on microlink_internal.h's ml_derp_conn_t, which doesn't have the entropy/ctr_drbg fields the source commit anchors near — this fork uses PSA crypto post-mbedTLS-4.x migration (same divergence already handled for issue #14 / PR #44). ml_derp.c and ml_wg_mgr.c applied cleanly against this fork's current tree; hand-ported all three files with the source commit's logic/comments unchanged.

• DERP TLS session resumption: save the negotiated mbedTLS session after each successful handshake, resume it on the next reconnect. Skips the full ECDHE handshake (~7.5s → sub-second) on reconnect-heavy/flaky networks. Best-effort — falls back to a full handshake transparently if resumption fails.
• DERP_CONNECT_TIMEOUT_MS 10s → 25s: the TLS handshake was observed taking ~7.5s+ and occasionally exceeding 10s on a lossy/high-latency captive network, tripping spurious "TLS handshake failed: timed out" → full-retry thrash.
• Per-peer exponential backoff on direct-path upgrade probes: on a P2P-hostile network (blocks peer-to-peer UDP) direct-path upgrades never succeed, so re-probing every 15s forever wastes CPU + DERP TX. Doubles the per-peer interval on each unanswered probe up to a 300s cap; resets to the base interval once a direct path is established.

FORK_PRS.md row 6 (issue #25) marked done.

Closes #25

Test plan

• idf.py build not runnable in this sandbox — needs a real ESP-IDF 6.x build to confirm.
• grep confirms all expected symbols (ML_DISCO_UPGRADE_BACKOFF_MAX_MS, upgrade_interval_ms, saved_session, have_saved_session, DERP_CONNECT_TIMEOUT_MS) present in the three touched files.
• Confirmed derp_free_tls_state() (the teardown/reconnect path) only frees ssl/ssl_conf, leaving saved_session intact across reconnects — matches intended lifetime.
• FORK_PRS.md diff touches only row 6; rows 7 (Fix wg_udp_output_cb thread-safety bug + RX queue starvation + disco_pong handshake gating (mined from antmanler/microlink) #26) and 9 (Investigate WG netif MTU 1420 vs 1280 (mined from antmanler/microlink) #28) remain open, deferred to a future session (bundled in a large, genuinely-diverged antmanler/microlink@2a7ba328 commit that needs surgical extraction).

🤖 Generated with Claude Code