fix: encode UNIXMS/UNIXMICRO times outside the UnixNano range - #826

With TimeFieldFormat set to TimeFormatUnixMs or TimeFormatUnixMicro, the JSON encoder computed the value as t.UnixNano()/div. UnixNano overflows for times before 1678 or after 2262, so the zero time.Time and far-future sentinels are logged as unrelated dates, while TimeFormatUnix prints the same values correctly:

UNIXMS before: {"zero":-6795364578871,"never":-4852116232933}
UNIXMS after:  {"zero":-62135596800000,"never":253402300799000}

AppendTime now uses t.UnixMilli() and t.UnixMicro(), and AppendTimes scales seconds and nanoseconds separately in the same way. UNIXNANO output is unchanged. One visible side effect: pre-1970 times with a sub-unit fraction now round down like TimeFormatUnix and UnixMilli (time.Unix(-1, 999999000) gives -1 ms instead of 0). I can drop that test case if you prefer a narrower change. #813 fixes the parsing side in ConsoleWriter and is independent of this one.

• Ran: go test ./..., go test -tags binary_log ./..., go vet ./... and gofmt -l on the changed files: all clean.
• Checked: the new TestAppendTimeUnixOutsideNanoRange fails on master with 10 assertion failures and passes with the fix.
• Checked: AppendTimes matches UnixMilli, UnixMicro and UnixNano for 200k random times between years 1 and 9999, and a 10-value UNIXMS benchmark shows no slowdown.