fix collector error handling

[kaysond]

Closes #944
Closes #945

Builds on the existing wip commit — master is merged in and the work is finished in commits on
top, so the diff of what changed since wip stays reviewable.

#944 — stop storing smartctl's failure output

smartctl.exit_status is now a typed bitmask
(webapp/backend/pkg/models/collector/smartctl_exit_status.go) matching
smartctl.h,
used by both collector.SmartInfo and models.Scan.

The bits split cleanly in two:

• bits 0-2 (FAILCMD, FAILDEV, FAILSMART) mean smartctl could not produce usable output —
  SmartctlExitStatus.IsFatal()
• bits 3-7 describe problems with the disk, which is exactly the data worth keeping

Two independent guards, because either side can be older than the other:

• the collector does not publish a result whose exit status is fatal
• UploadDeviceMetrics rejects such a payload with 400 before touching either database

LogSmartctlExitCode also became LogSmartctlExitStatus and now logs every set bit. The old
version was an if / else if chain, so exit status 0x18 only ever logged "failing disk" and
swallowed the pre-fail bit.

#942 (-n standby)

Exit status 2 is shared: it means both "device open failed" and "device is in a low power mode and
-n asked me to exit". When FAILDEV is the only bit set and the configured
metrics_smart_args contain -n/--nocheck, the drive is simply asleep — the collector logs that
at info level and moves on without publishing and without notifying.

That is deliberately conservative: a drive that has genuinely gone away also disappears from
smartctl --scan, so it never reaches Collect() in the first place.

#945 — notifications for collector failures

New route POST /api/collector/error, handled by handler.CollectorError. The collector reports:

• smartctl --scan failures (no device attached to the report)
• per-device smartctl --info failures — previously these were swallowed with //ignore errors.
  and the device silently vanished from the dashboard, because a device that fails --info never
  gets a ScrutinyUUID
• smartctl --xall failures that are fatal per the above

Detect.Start() is split into Scan() and Info(), and Info() returns a
[]detect.DeviceInfoError so each failing device is reported with its own error rather than one
shared message.

New setting metrics.notify_collector_errors, defaulting to true (opt-out, as the issue asked),
with a migration (m20260725120000) and a "Notify - Collector Errors" control in the settings dialog.

Differences from the wip commit

• POST /api/device/:scrutiny_uuid/smart had been dropped from server.go — restored. Without
  it the collector cannot upload anything.
• The lo.Filter predicate kept devices where ScrutinyUUID.IsNil() was true, i.e. exactly the
  devices that were meant to be dropped. Resolved in favour of the loop that landed on master in
  fix(collector): keep only devices with non-nil ScrutinyUUID #991.
• Two endpoints (/api/collector_scan_error and /api/device/:scrutiny_uuid/collector_error)
  collapsed into one. The device-scoped one had to invent a UUID via
  GenerateScrutinyUUID("", "", deviceName) for the --info failure case, which produces an ID that
  matches no real device; the single endpoint carries the device name and an optional UUID instead.
• SmartctlExitStatus lives in webapp/backend/pkg/models/collector rather than
  collector/pkg/models, so the dependency keeps pointing collector → webapp models (the direction
  detect.go already uses) instead of adding the reverse edge.
• Per-flag HasFailCmd()/HasFailDev()/… replaced by Has(flags), IsFatal() and
  Descriptions().

Tests

• smartctl_exit_status_test.go — which exit codes are fatal, and that all set bits get reported
• TestHasPowerModeCheck — the -n/--nocheck forms smartctl accepts
• TestUploadDeviceMetricsRoute_RejectsFatalSmartctlExitStatus — uploads good data, then the same
  payload with exit status 2, and asserts 400 plus that the stored summary is byte-for-byte unchanged
• TestCollectorErrorRoute — posts a collector error and asserts a notification actually reaches a
  webhook, with the right failure type and message

AI disclosure

Per AI_POLICY.md: the commits after wip were written by Claude Code (Opus 5).
It did the merge conflict resolution, the implementation, and the checks below, all inside containers
(golang:1.25-trixie with an influxdb:2.8 sidecar, node:24-trixie with headless Chromium);
nothing was installed on the host.

Automated verification performed:

• go build ./... and go vet clean; cross-compiled for windows/amd64, darwin/arm64 and
  freebsd/amd64 so the per-OS detect files all still compile
• go test ./... against a live influxdb — all packages pass, including the
  webapp/backend/pkg/web integration suite
• npm run build:prod and npx ng test --watch=false --browsers=ChromeHeadless --code-coverage
• gofmt clean

That is the extent of the verification behind this PR as opened. It has not been run against real
hardware, and in particular the -n standby path has not been exercised against a drive that is
actually asleep.