fix(mcp): synchronize async video job snapshots

[mekjr1]

Root cause

The async video worker updated mutable VideoJob fields under jobsMu, but get_flow_status released its read lock after retrieving the map pointer and then read those fields unlocked. Linux go test -race ./... caught the worker's status write racing the poller's status read.

Fix

• Treat jobsMu as the guard for both the jobs map and every mutable VideoJob field.
• Copy a coherent job snapshot while holding RLock, then build and marshal the poll response after unlocking.
• Snapshot generation inputs and publish processing/terminal transitions under short locks, with related terminal fields assigned atomically and Status assigned last.
• Keep generation, downloads, history writes, and filesystem work outside the lock.
• Strengthen the async MCP test with a channel-gated loopback daemon that proves submission returns while generation is blocked, observes processing, releases the backend, waits boundedly for completion, and verifies request/result/file payloads before cleanup.

Validation

• go test ./pkg/mcp -run '^TestAsyncVideoSubmitAndPoll$' -count=100
• go test ./pkg/mcp -count=20
• go vet ./...
• go build ./...
• git diff --check
• Local go test -race unavailable because this Windows Go environment has CGO_ENABLED=0; the CI Linux race job is authoritative.
• Local go test ./... reaches all packages but the unchanged pkg/gemini.TestRefreshSession fails because the installed Gemini Desktop app does not expose the required CDP endpoint on port 9223. The same normal suite passes on current-main Linux CI.