feat(alibaba): support wan3 all-in-one video generation (#19445)

Wan3 (wan3.0-video) is one model id that serves text-to-video,
image-to-video (first and last frame), and reference-to-video. The
existing provider derived both the wire protocol and the generation mode
from substrings of the model id (isWan27Model, detectMode), so wan3 fell
through both checks: it was classified t2v and sent the legacy wan2.6
shape — parameters.size instead of the resolution tier, no ratio, and
input.media never built, silently dropping every reference image and
video.

This splits protocol detection from mode detection. detectProtocol
returns legacy / wan27 / wan3; mode detection stays for the ids that
actually name their mode. wan3 builds input.media from whatever the
request carries, so a bare prompt is text-to-video, a start image is
image-to-video, and references are reference-to-video — with no id
suffix involved.

Also for wan3: last_frame is a real slot rather than an
ignored-with-warning option, the 480P tier is accepted alongside
720P/1080P, the audio boolean is forwarded (wan2.7 suppresses it because
audio is always on there), and shot_type warns as it does on wan2.7.
reference_audio, file, and link have no top-level call option, so
they're reachable through providerOptions.alibaba.media, whose type enum
is extended; ratio gains wan3's adaptive default.

doStatus now parses and surfaces wan3's extra usage facts —
input_video_duration, fps, ratio — as providerMetadata.alibaba.usage.
These are spread conditionally, so metadata for wan2.x is unchanged.
input_video_duration matters downstream: wan3 counts input video toward
its 30s ceiling, and consumers billing on duration need the output
figure separated from the total.

wan2.6/wan2.7 request shaping is untouched.