Pick the NVDEC surface from an explicit option, not from output_dtype

[NicolasHug]

The decode building block was asking for a 16-bit NVDEC surface by setting
output_dtype to FLOAT32, which nothing downstream of it reads: it does no
color conversion, so it has no output dtype. The only thing it wanted was
for the surface to keep every bit the source has.

Say that instead, with nvdec_surface_matches_source. The surface a full
decode picks is unchanged - it still follows output_dtype, because a frame
that is about to be converted to uint8 only needs an 8-bit surface.

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/meta-pytorch/torchcodec/1692

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[NicolasHug]

Not actually merged — nothing from this branch landed on main, which is still at 2312413.

This PR's base was packet-seam-const-avpacket. The stack was then reordered so that branch sits above nvdec-surface-depth instead of below it, which made this PR's head commit (56ee0a6) an ancestor of its own base branch. GitHub marks a PR merged when its head becomes reachable from its base, so it auto-closed on the force-push. The recorded "merge commit" is just this PR's own head SHA — there is no merge commit.

Since a merged PR can't be reopened, the change continues in #1696, which has the same commit with base main.