From<OutputItem> for Item (and InputItem) for multi-turn responses (#538)

* Add From<OutputItem> for Item / InputItem (multi-turn responses)

The Responses API requires Reasoning items to be echoed back into the
next request's input alongside the FunctionCall items they preceded.
Without that pairing the server rejects the request with:

    Item 'fc_...' of type 'function_call' was provided without its
    required 'reasoning' item: 'rs_...'

In the Python and Node SDKs the canonical pattern is `input_list +=
response.output` because their input/output item types are the same.
In Rust the schemas split into Item vs OutputItem, with several
variants that have *Resource (output, required id/status) and
*ItemParam (input, optional id/status) flavors, so callers had to
hand-write a per-variant conversion to do what's a one-liner in the
official SDKs.

Add From<OutputItem> for Item and From<OutputItem> for InputItem,
plus the supporting per-variant From impls (status enum mappings,
shell environment/outcome mappings, apply_patch operation mappings,
resource-to-param mappings). Required id/status fields fold into
Some(...) on the way in; the output-only created_by annotation is
dropped because the input schema doesn't accept it. The canonical
SDK pattern now works:

    input_items.extend(response.output.into_iter().map(InputItem::from));

Includes round-trip tests covering the headline reasoning +
function_call pairing case plus the resource-to-param folding for
ApplyPatch, Compaction, and FunctionCallOutput. New
responses-multi-turn-reasoning example demonstrates the pattern with
a reasoning model.

Closes #492.

* chore: fmt

Author: Madoshakalaka