feat(coap): per-type RPC dispatch over /zlet/<type>/# (#31)

[rodrigopex]

Summary

Phase 3 of the CoAP frontend adoption: per-type RPC dispatch over UDP.

• Codegen emits one COAP_RESOURCE_DEFINE per opted-in zephlet type at /zlet/<type>/# in <prefix>_coap_interface.c alongside a per-type POST handler. The handler parses instance + method out of the URI, validates z->api == &<type>_api defensively, and dispatches via a per-method switch with typed nanopb req/resp locals on the stack. The C compiler sizes the function frame from those locals — no cross-type max macros.
• The frontend gains shared zephlet_coap_send_{error,response} and URI-parsing helpers. zephlet_coap_frontend.c owns the single COAP_SERVICE_DEFINE(zlet_coap_service, …).
• Three new Kconfig knobs: MAX_SEGMENT_LEN, MAX_URI_SEGMENTS, CMD_TIMEOUT_MS. ZEPHLETS_COAP now selects COAP_EXTENDED_OPTIONS_LEN so URI segments up to the configured max survive coap_find_options without truncation.
• Functional test target vendors minimal tick + ui zephlets so the api-mismatch path runs end-to-end. Pytest cases cover happy + five unhappy paths plus the GET stub (4.05 until the events bridge lands). The test target uses NET_NATIVE_OFFLOADED_SOCKETS so the host pytest fixture reaches the guest CoAP server directly on 127.0.0.1 — no TAP setup, no root, no /dev/net/tun — and the same suite runs identically on bare Linux, Docker, and Colima.
• CI now also exercises modules/lib/zephlet/tests (the existing job only ran the app's per-zephlet integration tests). Codegen race fix added: defers add_dependencies(app <prefix>_codegen) via cmake_language(EVAL CODE …) so slower builders no longer compile main.c before the per-zephlet codegen finishes.

Test plan

• west twister --testsuite-root modules/lib/zephlet/tests -p native_sim --inline-logs (CI on Linux) — 35 of 35 test cases pass, 4 of 4 configurations green: coap.translate (ztest), shared.dispatch (ztest), coap_functional.smoke (pytest), coap_functional.rpc (pytest, 8 RPC cases).
• just docker-test from modules/lib/zephlet/ (local Apple Silicon, native_sim/native/64 via zephlet-tester:latest image) — same 35/35 pass.
• just c b (app build on mps2/an385) — green, footprint identical to pre-Phase-3 (FLASH 50556 B, RAM 13972 B).
• just test (app's per-zephlet integration tests) — 23/23 pass.
• west twister --testsuite-root modules/lib/zephlet/tests -p mps2/an385 (local) — 22/22 ztest cases pass; pytest configurations correctly filtered (mps2/an385 has no networking).

Notes

• The locked-decision "one catch-all /zlet/#" was relaxed in favor of per-type resources — each opted-in zephlet's codegen output owns its own resource registration. Gives each type self-contained CoAP routing and eliminates the catch-all's URI dispatch table.
• Per-method scratch lives on the dispatch function's stack as typed locals; the compiler computes the frame from real C declarations, no MAX_FROM_LIST aggregator or generated handler-max header.
• Phase 2's zephlet_coap_decode_request / ..._encode_response helpers stay in zephlet_coap_translate.c as a documented transport-free API for callers that prefer the wrapping; the dispatch path inlines pb_decode / pb_encode for explicitness.
• Dockerfile + justfile at the infra root give a one-shot just docker-build / just docker-test workflow for local iteration without re-installing Python deps per run.

Closes #31