fluent: fix panic on failing TLS connect

[mrkamel]

Fix for #119 supported by Claude Code

Problem

tls.DialWithDialer returns a concrete *tls.Conn, so assigning it
straight to f.conn stored a typed nil on failure: f.conn != nil
held while the pointer behind it was nil. Two nil derefs followed, both
reachable by pointing a tls logger at an endpoint that's down:

• close() called Close() on the nil *tls.Conn.
• syncConnectWithRetry skipped the reconnect, so syncWriteMessage
  dereferenced it.

Fix

Dial into a local in each branch and assign f.conn only after the dial
succeeds; on error connect() returns early without touching it. The
old implicit "clear f.conn on failure" was dead — all three callers
already enter with it nil. latestReconnectTime is unchanged. The
tcp/unix branches are retargeted for consistency only; they were
never affected.

Testing

TestNoPanicOnFailingTLSConnect points a sync TLS logger at a closed
loopback port and covers both crashes (two Posts, deferred Close).
It panics on the old code, passes with the fix. go test ./fluent/ -race -count=1 passes.

[thaJeztah]

cc @cosmo0920 ptal

[cosmo0920]

Thanks. It's reasonable change to me.

[thaJeztah]

Doh! Meant to comment on the PR, but did on the ticket 🙈

Thanks @cosmo0920 ! Any chance on a new patch release? Then we could be able to close out moby/moby#47709 on our side

I opened one minor PR to remove uses of the deprecated io/ioutil, and go fix, which should probably be safe to include as well;

• chore: go fix, remove deprecated io/ioutil, and fix error capitalization #145