ci: fix nuttx builds failing with 'no space left on device'

[matthargett]

Problem

compilation on nuttx is intermittently red on main and across PRs — every build_iwasm_on_nuttx leg fails with no space left on device, in two places: the container image pull (Docker pull failed, the dominant mode) and during the build while compiling the from-source bloaty (protobuf/abseil).

Root cause

It's a container: job on the large apache-nuttx-ci-linux image, and it also builds google/bloaty (recursive protobuf/abseil submodules) from source on every matrix leg for the size report. A container job pulls its image at job init, before any step runs, so no step can free disk first — and the runner's default free space plus that from-source build doesn't fit.

Fix

• Drop container:. Run on ubuntu-latest, remove the unused preinstalled toolchains (~25 GB) as step 1, then run the NuttX configure+build inside the apache image via docker run with the workspace mounted. The image is pulled after disk is freed (verified ~119 GB free), so both modes are fixed.
• Build bloaty once + cache it. A small build_bloaty job compiles the pinned bloaty in the image and caches the binary, keyed on BLOATY_REF — bump the ref and the cache key changes, so it rebuilds and refreshes automatically. The matrix needs: build_bloaty and restores the prebuilt binary, so 63 protobuf builds collapse to one. The size report (bloaty -d compileunits --source-filter wamr/libc-builtin/libc-wasi) is byte-for-byte unchanged — the MCU binary-size regression signal is fully preserved.
• Drop a duplicated wamr_config_option matrix entry (it was listed twice), removing a redundant leg.

Verification (on a fork)

• Free-disk gives /dev/root 145G 26G 119G 18% /; the apache image pulls; all board matrix legs build and link, no ENOSPC.
• build_bloaty builds once and Cache saved with key: bloaty-<ref>-ubuntu-latest; the matrix legs Cache restored from key: bloaty-<ref>-... (restore, not rebuild) and the size report prints.

CI cost

Measured on the fork runs: building bloaty (protobuf/abseil/…) from source adds ~3 min to every matrix leg (~8–9.5 min per leg vs ~5.7 min with a cache restore), and it was being done on all 56 legs. Doing it once (~7.6 min, and ~free on a warm cache) instead saves roughly 56 × 3 ≈ 170 Actions-minutes per run; dropping the duplicated matrix leg saves ~7 × 5.7 ≈ 40 more — about ~200 Actions-minutes per nuttx run.

At the current ~18 nuttx runs/week on this repo that's roughly ~3,600 Actions-minutes (~60 hours) per week (warm cache; a bloaty-ref bump adds one ~8-min rebuild that run). This is on top of no longer wasting minutes on the ENOSPC-failed runs.