Add PyDict_GetItem method to PyDict type binding

[MarshalX]

This will be helpful in micro optimizations where we don't care about errors and we are okay to get None.

The current implementation of get_item uses PyDict_GetItemWithError which creates a little overhead while taking Python errors.

pyo3/src/types/dict.rs

Lines 431 to 438 in dcba984

	match unsafe {
	ffi::PyDict_GetItemWithError(dict.as_ptr(), key.as_ptr())
	.assume_borrowed_or_opt(py)
	.map(Borrowed::to_owned)
	} {
	some @ Some(_) => Ok(some),
	None => PyErr::take(py).map(Err).transpose(),
	}

The overhead could be pretty displayed using flame graph:

The right side with PyErr::take could gone

The solution is to add new method get_item_TBD which will use PyDict_GetItem instead

[davidhewitt]

For some context, I changed this in #3330. I removed the "infallible" version but perhaps that was too aggressive and we can bring it back with an awkward name and a strong recommendation not to use it except in performance sensitive cases.

It's possible there are ways that we could make error handling overhead lighter, those would need exploration.

[davidhewitt]

Possibly relevant: in 3.13 we will have PyDict_GetItemRef which will be faster because error and not-present cases will be distinguishable in the return code. https://docs.python.org/dev/c-api/dict.html#c.PyDict_GetItemRef

So a cleaner option might be to just optimise this method on 3.13 once available.

[davidhewitt]

Also for freethreaded Python PyDict_GetItemRef is strongly recommended to avoid race conditions, which is a reason not to expose an API wrapping PyDict_GetItem as it is not thread safe.