Fix LanguageModel.create() not downloading model with multimodal

If the model is not downloaded yet, optimization guide does not know if
multimodal is supported. The check for capabilities should be moved
until after the download is complete.

This fixes the case where create({expectedInputs:[{type:"image"}]}) will
not download the model and fail immediately.

I looked into making a regression test for this but couldn't find a
good hook to test this change. I can look into that as a follow up.

Bug: 406326096
Change-Id: I77c4b2bf8bbd70ce59f022b9d88a403427308e00
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/6548077
Reviewed-by: Mike Wasserman <msw@chromium.org>
Commit-Queue: Clark DuVall <cduvall@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1460283}