[Platform] Accept Model instance in Platform::invoke()

[chr-hertel]

Q	A
Bug fix?	no
New feature?	yes
Docs?	yes
Issues	-
License	MIT

Platform bridges ship static model catalogs that go stale as providers release new models. Today, using a model that's not in the shipped catalog means registering additionalModels, swapping in a FallbackModelCatalog, or patching the lib.

This widens Platform::invoke() (and Provider::invoke() / Provider::supports()) to accept string|Model, so a fully defined model can be handed over per call, bypassing the catalog entirely:

use Symfony\AI\Platform\Bridge\OpenAi\Gpt;
use Symfony\AI\Platform\Capability;

$model = new Gpt('gpt-newest', [
    Capability::INPUT_MESSAGES,
    Capability::OUTPUT_TEXT,
    Capability::TOOL_CALLING,
], ['temperature' => 0.5]);

$result = $platform->invoke($model, $messages);

A string keeps the existing catalog-based routing untouched. A Model instance is routed to the first provider whose model clients accept it.

Note: you must pass a bridge-specific subclass (Gpt, Claude, …), not the base Model, since model clients, result converters and normalizers dispatch on the concrete class. All bridge model classes are directly instantiable (the final ones included — final only blocks subclassing, not new).