fix: timer metrics unit

[multani]

Histograms are built using the default bucket definitions:

If Buckets is left as nil or set to a slice of length zero, it is replaced by default buckets.

The default buckets are DefBuckets if no buckets for a native histogram (see below) are used, otherwise the default is no buckets.

And DefBuckets is:

var DefBuckets = [][float64](https://pkg.go.dev/builtin#float64){.005, .01, .025, .05, .1, .25, .5, 1, 2.5, 5, 10}

DefBuckets are the default Histogram buckets.

The default buckets are tailored to broadly measure the response time (in seconds) of a network service.

The timers were measuring values in milliseconds, whereas the underlying buckets are configured to store number as seconds.

I spotted the misconfiguration by checking the DNS resolution timers on our deployments: most of the timers are between 5 and 10 ms (to the last 2 buckets) and all the initial buckets are almost not used.

Changing to seconds instead of milliseconds would fill up in our case the first 2 or 3 buckets and show the higher latencies (if any) more correctly.

[cla-assistant]

All committers have signed the CLA.

[cla-assistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[multani]

I signed the CLA :)