[BUG] Precision issue using ts_format and ts_strftime in bloblang

I have encountered a precision issue using ts_format in bloblang mapping.

Input

root.time = (1555234838050/1000).ts_format("2006-01-02T15:04:05.999-07:00", "UTC")

Expected output

{
    "time": "2019-04-14T09:40:38.050+00:00"
}

Actual output

{
    "time": "2019-04-14T09:40:38.049+00:00"
}

Version

4.53.0

The same problem occurs using ts_strftime(). I read that these two methods support nanosecond precision, so ms precision shouldn't be a problem.
PS: I currently use a dirty trick to fix this problem and avoiding heavily string operation at the same time (which is critical for my high frequency use case):

root.time = (1555234838050/1000+0.0004).ts_format("2006-01-02T15:04:05.999-07:00", "UTC")

But I guess this can't be 100% safe as trying to build a feature on a bug.

Start by reproducing the issue's Bloblang mapping with ts_format and ts_strftime using the provided millisecond timestamp and compare the actual and expected outputs. Trace the implementation or entry points for these two methods; done means both preserve the expected millisecond precision without relying on the timestamp offset workaround.