fix(read): Duration() parses Go duration strings (e.g. "300s", "1h30m")

[j-ibarra]

The Config.Duration getter reads the value as an int64 and returns time.Duration(n) — i.e. nanoseconds — so:

• a duration string like "300s" cannot be read at all: it fails integer parsing and yields 0 (not even the default);
• a bare number like 300 is interpreted as 300ns.

Duration-string parsing existed only on the struct-binding path (the ParseTime decode hook via reflects.ToTimeOrDuration), never on the Duration() getter — so config.Duration("ttl") on a ttl: 300s value silently returns the wrong result.

This parses the stored value as a Go duration string first (time.ParseDuration), falling back to the previous int64-nanoseconds behaviour for bare integers, then to the default:

func (c *Config) Duration(key string, defVal ...time.Duration) time.Duration {
	str, exist := c.getString(key)
	if !exist {
		if len(defVal) > 0 { return defVal[0] }
		return 0
	}
	if dur, err := time.ParseDuration(str); err == nil { // "300s", "1h30m", ...
		return dur
	}
	if n, err := strconv.ParseInt(str, 10, 64); err == nil { // bare int = nanoseconds (unchanged)
		return time.Duration(n)
	}
	if len(defVal) > 0 { return defVal[0] }
	return 0
}

Backwards compatible: bare integers keep their nanosecond meaning; only previously-unreadable duration strings change (from 0/error to the parsed value). No existing test asserted the old getter behaviour (the existing TestDuration covers the Decode/ParseTime path). Adds a getter test covering "300s"/"1h30m"/"20m", a bare integer, and the malformed/missing default fallbacks. Full package test suite passes.