Initialize MEOS per worker thread

[estebanzimanyi]

The MEOS 1.4 uplift keeps the session timezone, errno, PROJ context and RNGs in thread-local storage and requires every thread that calls into MEOS to run meos_initialize() before its first call. The extension only initialized MEOS once on the load thread via std::call_once, so DuckDB TaskScheduler worker threads executed scalar, cast and aggregate bodies with a NULL session_timezone and segfaulted in pg_next_dst_boundary on the first timestamp parse (reproduced locally as a SIGSEGV in 042_tgeogpoint_parity, previously masked because amd64 CI never reached the tests and arm64 runs no test step). A thread-local guard now runs the per-thread init and re-installs the process-global MobilityDuck error handler (meos_initialize() resets it to the exit-on-error default), invoked from the scalar exec wrapper and through a cast registration trampoline that covers every cast entry point. Locally the full suite goes from unrunnable to 58/59 files and 1311/1312 assertions with no segfaults; the one remaining assertion is an unrelated pre-existing ln() output drift from the MEOS bump.

[estebanzimanyi]

Reviewer's quickstart — ~8 minutes

What this PR does in one sentence: moves meos_initialize() from process-once to per-DuckDB-worker-thread, fixing the data-race where MEOS's thread-local state (timezone, error handler, lwgeom session) was uninitialised in workers spawned after the initial LoadInternal.

Files (14): mostly src/geo/*.cpp and src/temporal/*.cpp — each scalar/cast/cast-trampoline now starts with MeosThread::ensureReady() (a TLS-guarded one-time init).

Read first:

• src/include/mobilityduck/meos_thread.hpp (new) — the TLS guard + ensureReady() definition.
• src/include/mobilityduck/meos_exec_serial.hpp (new) — the serial-execution wrapper for paths MEOS itself isn't thread-safe on (pg-derived TZ, legacy GEOS).
• src/mobilityduck_extension.cpp — registers the worker thread's MEOS state.
• 11 callsite files just add MeosThread::ensureReady(); at the top of each UDF lambda body.

Verification:

# Repro pre-PR: SIGSEGV on multi-threaded queries
duckdb -unsigned -c "LOAD mobilityduck; SET threads=4; SELECT count(*) FROM (SELECT atTime(traj::VARCHAR, '2026-01-01'::TIMESTAMP) FROM ...);"
# Pre-PR: occasionally crashes. Post-PR: clean.

Cross-link: Linux arm64 needs #161 for the MeosType build error.

Why it's safe to merge: pure additive guards; valid-input paths are unaffected. The change correctly diagnoses MEOS's process-singleton initialization model as a multi-worker bug.

[estebanzimanyi]

Folded into #134: the MEOS pin bump and the changes that keep CI green (graceful icu degradation + stage_icu, per-thread MEOS init, deterministic ln/exp/log10, macOS int64 forwarder, wasm pg_config) are jointly required for the extension to build, load, and pass — they ship as one foundation PR.