BUG: keep value types when calling .items() on Series

[ghost]

• closes BUG: to_dict() unexpectedly casts np.float32 to python float #50125
• closes Series iteration and to_dict methods *sometimes* return underlying storage type vs. Python object #25969
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.

I have identified that the .items() method of Series modifies the type of values, before returning them.
One of the major consequences (for me) is that the .to_dict() method unexpectedly returns modified types.
This can be problematic with float32 or float16 types, as .items() would convert them to Python's float, thus changing the actual value of the output by adding some extra decimal digits.

.items() modifies the type because it calls the .__iter__() method of the parent's class base.IndexOpsMixin, which enforces the return of Python types.

From my point, .items() and .to_dict() should not change the type of the values. This can be easily done by refering to Series._values, as proposed in this PR.

[ghost]

Perhaps I jumped the gun on this one, as it could be intentional to return python type instead of original type? : #20791 (comment)

I obviously don't have an holistic view on this one, but my user experience was not optimal, when I got to_dict() results different from what I would see in the original DF.

[phofl]

We want to return python types in I/0

[ghost]

We want to return python types in I/0

I see. Is there anything we can do to improve the conversion to float?
My initial issue came from using .to_dict() after .round(), which under the hood does something like:

np.float32(123.456789).round(3).item()
# 123.45700073242188 when expecting 123.457

Hence I see rounded values all around my work on pandas DF, but when I finally output a dict to feed an API, it has unexpected decimals.

[phofl]

You'll have to check where things go wrong to see if we can improve. Not familiar with the implementation myself

[ghost]

OK, it seems like this is going to be unavoidable.
When converting from np.float32 to python's float, I don't see a way to do that more cleanly than it is.
So be it, if to_dict() must return python types, then we have to live with this kind of trouble.