#2023 Video input: bounded preparation progress and cancellation follow-up

Beschreibung

## Summary

Follow-up to #1609, based on v0.1.43. The six original #1609 commits are preserved as individual cherry-picks with their original authorship; this branch adds the serving/lifecycle work needed for long video preparation.

- report bounded `VPROG <frames_done> <frames_total>` updates from `strata-vision` once per 64-frame batch;
- poll encoder replies instead of blocking for the whole video, with validation and the existing timeout;
- bound concurrent video preparation by the engine batch capacity and expose queue/preparation/progress state in `/status` and `/metrics`;
- cancel queued and in-progress preparation when a client disconnects, kill/wait for the encoder, remove partial files, and restart the encoder on the next request;
- propagate cancellation through Chat Completions, Responses, Anthropic, and MCP preparation paths;
- show preparation stage and bounded progress in the web UI.

The follow-up does not combine the incompatible SVE2 video implementation from #1327.

## Validation

- clean isolated CUDA engine and vision builds under `/tmp/strata-video-followup-build`;
- video/progress/cancellation tests, `ClientHangUp`, and the isolated Responses sampler test pass;
- `gguf_split_test` and `gguf_reader_test` pass;
- the combined `serve.test_server serve.test_frontend serve.test_responses serve.test_parallel` run covered 388 tests with one skip. The existing order-sensitive `NoLeakedSampler.test_server_close_ends_the_sampler` assertion passed in one run and failed in a later full-order rerun, but passes alone; it is unrelated to this video change;
- live AP-Q5 CUDA smoke test on RTX 3090 + RTX 5080: OpenAI Chat Completions video returned HTTP 200, Responses `input_video` returned HTTP 200, `/metrics` observed `preparing_video` and `VPROG 64/122`, client-disconnect cancellation returned to idle, and the next video request succeeded after encoder restart. GPU 2's unrelated workload was not disturbed.