refactor(get): move the get command's body into internal/get

[dimoschi]

Closes #25.

What

cmd/get.go held the PoolManager construction and the json/text/spew output switch inside a cobra Run closure, none of it reachable from a test. That is the shape server.go had before #20, and it is why the mutation survivors in #24 had to be accepted rather than covered.

internal/get now exposes Run(cfg, w io.Writer) error, writing to an injected writer so every format can be asserted. cmd/get.go is a Config, a call, and an exit code.

Chasing the survivors found three real defects

Rather than accept what the mutation gate reported, each survivor was treated as either dead code or a missing test. Three turned out to be dead or wrong code:

The JSON encode error was unreachable. Everything PoolManager marshals is a string, int64, map or slice, and both custom MarshalJSON methods return a nil error unconditionally, so json.Marshal cannot fail. Encoding straight to the writer removes the branch and leaves one error a caller can actually hit — the write.

uitable.Wrap was dead. Rendering the same table with it on and off is byte-identical, because MaxColWidth elides the cell first. Deleted rather than pinned. The width itself does matter and is now asserted through its effect, by measuring the widest rendered cell.

Every failed scrape was logged twice. Pool.error logged it and collectPools logged it again with context, so the server path emitted two identical-ish lines per failure per scrape. That predates this branch; returning the joined error from Update in #24 had quietly made it three. Pool.error now only records, leaving the line that carries context.

And two broken tests of my own

Worth stating, because both passed while testing nothing:

• The test logger discarded Info, Debug and Error, keeping only the *f variants. The duplicate-logging test therefore passed against the exact duplication it was written to catch.
• The column-width assertion compared the measured cell against maxColWidth, so mutating the constant moved both sides together. It now compares against 80 directly and fails on an off-by-one in either direction, which I verified by making the change and watching it fail.

Behaviour changes

• An unrecognised --out value is an error rather than a logged message, so it exits non-zero instead of silently succeeding with no output.
• Writing to stdout is checked, so a failed write is reported rather than discarded.
• --out json gains a trailing newline, a side effect of streaming through json.Encoder. Parsers do not care; jq pipelines are unaffected.

Verification

$ mise exec -- go test ./... -race
ok  	github.com/hackthebox/php-fpm_exporter/internal/get
ok  	github.com/hackthebox/php-fpm_exporter/internal/server
ok  	github.com/hackthebox/php-fpm_exporter/phpfpm
ok  	github.com/hackthebox/php-fpm_exporter/phpfpm/fcgi

All three output formats checked against the built binary, plus the exit codes (1 on scrape failure, 0 otherwise).

golangci-lint (v2.12.2, as CI pins): 4 issues, unchanged from master, none new.

Mutation survivors went 23 → 7. The remaining 7 are all in cmd/get.go — the Config literal, the if err != nil, and the log.Error — and are accepted because covering them means putting tests back into cmd, which is what this PR exists to avoid. Nothing in internal/ or phpfpm/ survives.