stop dying when a container disappears mid-stats

[mrq1911]

Fixes swarmpit/swarmpit#736.

Root cause

ContainerUsage declares var v *types.StatsJSON and then:

dec := json.NewDecoder(resp.Body)
if err := dec.Decode(&v); err != nil {
    dec = json.NewDecoder(io.MultiReader(dec.Buffered(), resp.Body))
}          // rebuilds the decoder, never retries, never returns
...
previousCPU = v.PreCPUStats.CPUUsage.TotalUsage    // stats.go:157 - v is still nil

The error branch is a no-op left over from Docker CLI's stats loop, which retries — here there is no loop, so v stays nil and the very next line dereferences it. stats.go:157 is exactly the frame in the reported trace.

null is the nastier case: it decodes successfully and leaves v nil, so an error check alone isn't enough.

Verified against the pre-fix logic — all four bodies panic with the reported error:

body=""                   -> invalid memory address or nil pointer dereference
body="null"               -> invalid memory address or nil pointer dereference
body="{\"cpu_stats\": "   -> invalid memory address or nil pointer dereference
body="<html>error</html>" -> invalid memory address or nil pointer dereference

This happens routinely: a container that exits while its stats are being read returns an empty or truncated body.

Why it takes the whole agent down

ContainerUsage runs in a goroutine (ContainersUsage, stats.go:122) and there is no recover() anywhere in the agent, so one dying container panics the process. That makes it a second, independent cause of "Statistics not ready" in the UI — the agent logs Stats collector started, then goes silent.

Changes

• return early when the body doesn't decode, or decodes to nil
• ContainerUsage now returns (status, ok) so failed reads aren't reported as zero-valued containers
• recover() per collection goroutine, so a future panic in one container degrades that container instead of killing the agent
• drop the now-unused io import

Tests

New swarmpit/task/stats_test.go drives ContainerUsage/ContainersUsage against an httptest fake of the daemon stats endpoint: empty, null, truncated, non-object and HTML bodies all return ok=false instead of panicking; a well-formed body still parses; and a mixed run reports the healthy container while skipping the bad one.

go build ./..., go vet ./... and go test ./swarmpit/task/... pass on golang:1.12 (matching the Dockerfile). Only stats.go and the new test are touched — go.mod is unchanged, and I left the repo's pre-existing gofmt drift in other files alone.