Fix incorrect cast Timestamp with Timezone

[aprimadi]

Which issue does this PR close?

Closes apache/datafusion#5952.
Closes #1936

Rationale for this change

N/A

What changes are included in this PR?

N/A

Are there any user-facing changes?

No

[tustvold]

Thank you for this, I'm not sure this logic is correct, however.

If both quantities have timezones, no adjustment should be necessary, as they are already normalized to UTC.

The complexity arises when converting between timestamps where one doesn't have a timezone. This requires some care, especially to handle timezones that don't have fixed offsets.

In the forward direction I think this involves doing something like

let from_tz = from_tz.parse()?;
array.unary_opt(|i| {
    T::make_value(as_datetime_with_timezone(i, from_tz)?.naive_local())
})

And in the reverse direction

let to_tz = to_tz.parse()?;
array.unary_opt(|i| {
    let local = as_datetime(i);
    let offset = to_tz.offset_from_local_datetime(&local).single()?;
    T::make_value(local + offset.fix())
})

Let me know if you get stuck, timezones are wild...

[aprimadi]

@tustvold Yes sorry it's not done yet. I have problem compiling in my macbook pro (see Discord). That's why I checked in the code that isn't even compiling yet so I can switch to my Linux machine.

[aprimadi]

Thank you for the feedback @tustvold. Very helpful.

[tustvold]

This is great, really awesome work

None conversion

[tustvold]

I think we just need to move the tests and this is ready to go, returning UTC time when stripping the timezone is what we currently do, and is consistent with the parsing kernels, so 👍 from me

[tustvold]

Thanks again 👍

[aprimadi]

Thanks for the quick review @tustvold and very helpful feedback along the way. 🍻

[alamb]

See related discussion #5827