fix: make the daemon fail fast and put it under a systemd watchdog

[richardkiene]

Why

The daemon swallowed every error: an unparseable config.json was treated as "all GPUs auto", a missing or corrupt curve.json fell back to a hardcoded default, an unknown mode fell back to that default too, and a failed nvmlDeviceSetFanSpeed_v2 was ignored — the service sat active while writing to stderr every 5 s. A daemon that hung inside NVML was never restarted, which in curve mode leaves the fan pinned at whatever speed was last set (possibly the 30% floor) while load arrives.

What

Daemon

• Every error is fatal: reason to syslog, fans reset to auto, exit non-zero so Restart=on-failure takes over. Covers unreadable config/curve, manual mode without a valid speed, unknown mode, temperature read failure, and any fan command NVML rejects.
• curve.json is re-read every poll (as config.json already was), so nvfd curve <temp> <speed> takes effect within 5 s without a reload. SIGHUP is kept as a logged no-op so a stray HUP does not kill the daemon; ExecReload is removed and the README updated.
• sd_notify READY=1 / WATCHDOG=1 / STOPPING=1 through a 40-line AF_UNIX datagram shim (src/notify.c) — no libsystemd link dependency. Unit is Type=notify, WatchdogSec=30 against the 5 s poll; the daemon refuses to start if WATCHDOG_USEC is under twice the poll interval.
• First poll hands every GPU not in manual/curve mode back to the driver, so an instance that died uncleanly cannot leave those fans pinned either.
• StartLimitIntervalSec=60 / StartLimitBurst=5 so persistent failure parks the unit in failed instead of restarting forever.

Readers

• config_read() distinguishes "no file yet" (empty object) from "file exists but unusable" (NULL + config_last_error()). config_write_gpu refuses to overwrite a config it could not parse. JSON_REJECT_DUPLICATES on both files.
• curve_load() validates: top-level object, integer keys 0–100 with a leading digit, integer values 0–100, no duplicate temperatures ("30" and "030" divided by zero in curve_interpolate), at most MAX_CURVE_POINTS, at least one point. curve_require() is the shared "load or explain why not" for callers that cannot proceed without a curve.
• Legacy plain-text migration refuses values the daemon would reject instead of writing them; a migration failure is fatal.
• fan_get_count returns -1 on error so fan_reset_to_auto cannot reset zero fans and report success.

CLI / TUI

• nvfd curve and nvfd <n> curve verify the curve loads before switching a GPU into curve mode; the TUI does the same before writing "curve", and exits with the reason rather than falling back to a built-in curve the daemon does not have. curve_default_interpolate is removed.
• Non-zero exit status when a command fails.
• The TUI and nvfd status exit with the parse error instead of rendering a blank config; the curve editor refuses to open on an invalid file rather than overwriting it from the default on save.