Make seconv conversion cancellation-aware and serialize global state

[Blackspirits]

Summary

This is prerequisite hardening for long-lived/concurrent callers such as the proposed seconv mcp server.

SubtitleConverter currently mutates process-wide state during a run:

• Configuration.Settings.General.CurrentFrameRate while loading frame-based formats;
• static SpellCheckConfig delegates for Fix Common Errors/OCR;
• translation prompt/model settings through AutoTranslateRunner.

A CLI invocation normally has only one conversion, but a long-lived server can invoke multiple converters concurrently and cross-contaminate those globals.

The converter also had no effective cancellation contract: Spectre passes a CancellationToken to ConvertCommand, but the command discarded it, and translation was explicitly invoked with CancellationToken.None.

This change:

• serializes SubtitleConverter runs behind a process-wide async gate while conversion still depends on mutable global libse state;
• adds an optional CancellationToken to ConvertAsync without breaking existing callers;
• passes the CLI command token into the converter;
• propagates cancellation through files, tracks, image-preservation stages, Task.Run, and auto-translation;
• rethrows caller cancellation at every conversion error boundary rather than recording it as an ordinary failed file;
• adds checkpoints around synchronous native/bitmap stages that cannot themselves accept a token;
• adds a regression test proving a pre-cancelled conversion throws before writing output.

Scope / limitation

Cancellation is cooperative. Synchronous legacy/native helpers without a token cannot be pre-empted inside the call; cancellation is observed before those stages and between work units. A completed synchronous write is not retroactively reported as cancelled, avoiding retry ambiguity after side effects already exist. The serialization gate is intentionally conservative until the remaining process-global settings become per-conversion state.

Validation

• based on upstream c77c4b428649ce7030e6cf218b4d7c384b955232
• changed files: SubtitleConverter.cs, ConvertCommand.cs, one focused cancellation test
• final HEAD: f83e30793665214f054f411feb0c2254b8b68d8a
• restore and build passed
• SeConvTests passed on the first run: 476 passed, 2 skipped, 0 failed (478 total)
• LibUiLogicTests passed on the first run: 873 passed, 0 skipped, 0 failed
• LibSETests passed on the first run: 2,008 passed, 0 skipped, 0 failed
• the first UITests run hit one unrelated UI/keymap flake: NativeKeymapShiftDeleteCutTests.Delete_WithoutShift_StillJustDeletes_InTheSyntaxTextEditor
• the workflow retry in a fresh process passed the complete suite; UITests: 5,151 passed, 9 skipped, 0 failed (5,160 total)
• retry was used only because of that unrelated UI flake
• final validation run: #34791238162
• audit draft only; no merge/promotion intended

AI assistance: ChatGPT was used to audit seconv cancellation propagation and concurrency hazards exposed by long-lived MCP callers.

[Blackspirits]

Independent adversarial re-check: conversion remains backwards-compatible through an optional CancellationToken, the CLI now forwards its token, caller cancellation is rethrown through every converter error boundary, and conversion runs are serialized while libse/seconv still mutate process-global settings. Cancellation around synchronous legacy/native work is cooperative rather than pre-emptive, which is documented in the PR. Full CI #34788562083 passed restore, build and the complete solution suite on the first run; SeConvTests reported 476 passed, 2 skipped and 0 failed, and UITests 5,151 passed, 9 skipped and 0 failed. Retry was not used. No blocker identified. Keep draft; no merge performed.

[Blackspirits]

Final adversarial re-check on f83e307: cancellation is propagated through the converter and caller cancellation escapes every conversion error boundary. The final correction also avoids checking for cancellation after a completed synchronous write/extraction, so an already-produced output is not reported as cancelled and accidentally retried. CI #34791238162 passed restore/build. SeConvTests (476/2/0), LibUiLogicTests (873/0/0) and LibSETests (2008/0/0) passed on the first run. The first UITests run had one failure in NativeKeymapShiftDeleteCutTests.Delete_WithoutShift_StillJustDeletes_InTheSyntaxTextEditor, a UI/keymap test outside all three changed seconv files; the fresh-process retry passed UITests 5151/9/0 and the entire suite. This is consistent with an unrelated UI flake, not a #109 regression. No blocker identified. Keep draft; no merge performed.