Fix POSIXt precision and tzone

[r2evans]

Two steps here for POSIXt value preservation:

• sets "tzone" attribute when missing/empty
• add options(bigrquery.digits.secs=6) to allow sending up to 6 sub-second digits to POSIXt-class objects

Before this patch,

bq_auth(path="my_token.json")
tbl <- data.frame(dbl=pi, psx=Sys.time())
ds <- bigrquery::bq_table(...)
bq_table_upload(ds, tbl); tbl2 <- bq_table_download(ds); bq_table_delete(ds)
tbl$psx - tbl2$psx
# Time difference of 5.000016 hours

Just fixing "tzone", we can get

bq_table_upload(ds, tbl); tbl3 <- bq_table_download(ds); bq_table_delete(ds)
tbl$psx - tbl3$psx
# Time difference of 0.0591259 secs

To support microsecond precision,

options(bigrquery.digits=22, bigrquery.digits.secs=6)
bq_table_upload(ds, tbl); tbl4 <- bq_table_download(ds); bq_table_delete(ds)
tbl$psx - tbl4$psx
# Time difference of 9.536743e-07 secs

Includes NEWS, code, and unit-tests.

[hadley]

Do you think we need to make this an option? i.e. why not always use 6 digits?

[r2evans]

Fair question. Two reasons, for better or worse:

1. Backward compatibility, so that revdeps don't see starkly-changing behavior.
2. Consistency with jsonlite's notion of "smaller is better" (regardless of lossy-ness).

Some (perhaps even most) people may not care about millisecond-resolution, I don't assume everybody is as concerned about it as I am. If it's hard-coded 6 then I'll be fine.

[hadley]

I think I'd prefer making it a constant. I don't think it should negatively affect revdeps; it might change some results in very small ways, but those changes should generally be for the better.

[hadley]

Big picture now looks good, so it's time for the nit-picking 😄 Thanks for working on this!